VISVESVARAYA TECHNOLOGICAL UNIVERSITY

“Jnana Sangama”, Belagavi-560018.

1st PHASE PROJECT REPORTATION

ON

“HUMAN RESOURCE MANAGEMENT SYSTEM WITH AI”

Submitted in partial fulfillment for the award of requirements for VII semester
Bachelor of Engineering in Computer Science and Engineering
By

J BHAVANI - 1CK20CS023
JYOTHI V- 1CK20CS024
SUCHITHA J - 1CK20CS050
THRIPURASRI S - 1CK20CS057
Under the Guidance of
Dr VANDANA DUBEY
Professor, Dept.of CSE.

C.BYREGOWDA INSTITUTE OF TECHNOLOGY

An ISO 9001:2015 Certified Institute
DEPARTMENT OF COMPUTER SCIENCE & ENGINEERING
Kolar-Srinivaspur Road, Kolar – 563101
2023-24
 C. BYREGOWDA INSTITUTE OF TECHNOLOGY
DEPARTMENT OF COMPUTER SCIENCE & ENGINEERING
(Srinivaspur Road, Kolar (T & D)
Kolar-563101)

CERTIFICATE
This is to certify that the 1st Phase project work entitled “HUMAN RESOURCE
MANAGEMENT SYSTEM WITH AI ”is a Bonafide work carried out J BHAVANI (USN:
1CK20CS023), JYOTHI V(USN:1CK20CS024), SUCHITHA J(USN:1CK20CS050) and
THRIPURASRI S(USN:1CK20CS057) in partial fulfillment for the award of Bachelor of
Engineering in Computer Science & Engineering of the Visvesvaraya Technological
University, Belagavi during the year 2022-23. It is certified that all corrections/suggestions
indicated for the internal assessment have been incorporated in the report. The project report
has been approved as it satisfies the academic requirements in respect of project work
prescribed for the VII Semester Bachelor of Engineering Degree.

_________________ ______________________ __________________

Signature of Guide Signature of HOD Signature of Principal
Dr. Vandana Dubey Dr. S N Chandrashekara Dr. Aprameyan S
Professor Prof & HOD, Principal,
Dept, of CSE, CBIT Dept, of CSE, CBIT CBIT-Kolar
Kolar Kolar
 ABSTRACT
In this Flask-based web application, we have developed a robust system that addresses key
challenges in the hiring process and employee management. Leveraging the Resume parse
library, the platform automates the analysis of resumes by extracting pertinent skills and
qualifications. This functionality proves invaluable in the initial stages of recruitment, allowing
for the swift and accurate comparison of candidate profiles against job requirements. By
automating this critical step, the application significantly streamlines the shortlisting process,
saving time and reducing the likelihood of overlooking qualified candidates.

Furthermore, our system goes beyond traditional hiring tools by incorporating a sophisticated
face recognition system for automating employee attendance tracking. This feature enhances
workforce management by eliminating the need for manual attendance recording, ensuring
accuracy, and saving administrative time. The integration of face recognition technology adds
a layer of efficiency and security to the attendance tracking process, providing a seamless
experience for both employees and administrators.

In essence, our Flask-based web application offers a comprehensive solution that synergizes
talent acquisition and workforce management. The combination of automated resume analysis
and face recognition attendance tracking not only reduces manual effort but also enhances
overall efficiency. This innovative platform marks a significant advancement in the realm of
HR technology, empowering organizations to make data-driven decisions and optimize their
hiring processes while ensuring accurate and automated attendance management.

I
 DECLARATION
We, J BHAVANI bearing USN:1CK20CS023, JYOTHI V bearing USN:1CK20CS024,
SUCHITHA J bearing USN: 1CK20CS050 and THRIPURASRI S bearing USN:
1CK20CS057 students of 7th semester B.E, Computer Science and Engineering of VTU,
declare that this project report entitled “HUMAN RESOURCE MANAGEMENT SYSTEM
WITH AI” , embodies the report of the project work carried out under the guidance of Dr
VANADANA DUBEY, Dept. of CSE, CBIT as partial fulfillment of the requirement of the
award of the degree in Bachelor of Engineering, Computer Science and Engineering, affiliated
to VISVESVARAYA TECHNOLOGICAL UNIVERSITY,BELAGAVI during the
academic year 2023-24. Further the content embodies in the project has not been submitted
previously by anybody for the award of any other degree.

Place: Kolar Signature of the Students

Date:

(J BHAVANI)

(JYOTHI V)

(SUCHITHA J)

(THRIPURASRI S)

II
 ACKNOWLEDGEMENT
The completion of any work is a showcase of constant dedication and co-operation of
many people who lent their hands which went seen or unseen.

We are grateful to my Institution and Management, C. BYREGOWDA INSTITUTE

OF TECHNOLOGY with its ideals and inspirations for having provided us with the facilities,
which has made this, project a success.

We would like to thank our beloved Principal, Dr. APRAMEYAN S CBIT for his kind
disposition for completing this undertaking successfully.

We express my heart full gratitude to Dr. S N CHANDRASHEKARA, Professor and

HOD, Department of Computer science & Engineering, CBIT, for giving me guidance,
valuable advice and support.

We extend our gratitude to our Project co-ordinator Prof. VASUDEVA R, Assistant

Professor, Department of Computer Science & Engineering, CBIT for her valuable advice,
support and constructive suggestions.

We extend our gratitude to our Guide Dr VANDANA DUBEY, Professor, Department

of Computer Science & Engineering, CBIT for valuable advice, support and constructive
suggestion.

We also thank to all our professors and the entire Department of Computer Science &
Engineering for their co-operation and suggestions.

The report would be incomplete if our do not thank my parents and friends for their
continuous encouragement and moral support.

J BHAVANI[1CK20CS023]

JYOTHI V[1CK20CS024]

SUCHITHA J[1CK20CS050]

THRIPURASRI S[1CK20CS057]

III
 TABLE OF CONTENTS

CHAPTER PARTICULARS PAGENO

NO
Abstract I

Declaration II

Acknowledgement III

1 INTRODUCTION 1
1.1 DEFINITIONS 1

1.2 PROBLEM STATEMENT 2

1.3 SCOPE OF PROJECT 2

1.4 OBJECTIVES 3

1.5 ORGANIZATION OF THE 3

PROJECT
2 LITERATURE SURVEY 4
2.1 RELATED WORK 4

3 SYSTEM REQUIREMENT 9
SPECIFICATION
3.1 HARDWARE REQUIREMENTS 9
3.2 SOFTWARE REQUIREMENTS 9
3.3 FUNCTIONAL REQUIREMENTS 10
3.4 NON-FUNCTIONAL 11
REQUIREMENTS
4 SYSTEM ANALYSIS 12
4.1 EXISTING SYSTEM 12

4.2 PROPOSED SYSTEM 12

IV
IV
CHAPTER 1

INTRODUCTION

1.1 DEFINITIONS
In the rapidly evolving landscape of talent acquisition and workforce management,
businesses are constantly seeking innovative solutions to streamline their processes and
enhance overall efficiency. The introduction of a Flask-based web application represents a
significant leap forward in this endeavour, offering a comprehensive platform that seamlessly
integrates multiple functionalities to address the challenges associated with hiring and
employee management. This dynamic system leverages the power of the Resumeparse library,
a sophisticated tool designed for resume analysis, to automatically extract and evaluate crucial
information from job applicants' resumes.
At the heart of this application is its ability to meticulously analyse resumes,
focusing on extracting relevant skills and qualifications. This automated process significantly
expedites the traditionally time-consuming task of resume screening. By comparing the
extracted skills with predefined job requirements, the system facilitates the swift shortlisting
of resumes, ensuring that only the most qualified candidates progress to subsequent stages of
the hiring process. This feature not only saves valuable time for recruiters but also enhances
the accuracy and consistency of candidate evaluation.
Beyond revolutionizing the hiring process, this Flask-based application goes a step
further by incorporating a cutting-edge face recognition system for automated employee
attendance tracking. This forward-looking technology eliminates the need for manual
attendance registers or time-tracking systems, providing a seamless and efficient solution for
managing workforce attendance. The integration of such a feature not only reduces
administrative overhead but also ensures accurate and real-time attendance data, contributing
to a more streamlined and accountable workforce management process.
By combining these advanced features, the platform serves as a one-stop solution
for talent acquisition and workforce management. The synergistic blend of automated resume
parsing and face recognition technology establishes a powerful tool that not only meets the
demands of modern recruitment but also addresses the evolving needs of businesses for
efficient and intelligent employee management. In essence, this Flask-based web application
represents a significant stride towards a more technologically advanced and automated future

1
HRMS WITH AI INTRODUCTION

for human resources, promising to redefine the standards of efficiency and effectiveness in
hiring and attendance tracking processes.

1.2 PROBLEM STATEMENT

The problem we have come across is the partial decisions from superiors which
suppress the talents in the industry, elongated recruitment process and managing the large
employee information and there is no secured system to monitor the employee workflow.
The current hiring and employee management processes are often labour-intensive,
time-consuming, and prone to human error. Identifying qualified candidates from a pool of
resumes can be challenging, and manual attendance tracking is both inefficient and susceptible
to inaccuracies. This Flask-based web application aims to address these challenges by
leveraging the Resumeparse library for automated resume analysis, skill extraction, and
comparison with job requirements. The integration of a face recognition system further
streamlines the attendance tracking process. The goal is to create a robust platform that
significantly reduces the manual effort involved in talent acquisition and workforce
management, ultimately enhancing efficiency, minimizing errors, and providing a
comprehensive solution for organizations seeking to optimize their hiring and attendance
processes.

1.3 SCOPE OF THE PROJECT

The future scope of this Flask-based web application is promising, with
opportunities to further enhance its capabilities and address evolving needs in talent acquisition
and workforce management. Future iterations could incorporate machine learning algorithms
to continuously improve and refine the resume analysis process, enabling the system to adapt
to changing job market trends and skill requirements. Integration with additional third-party
tools and platforms could expand the application's functionalities, such as integrating with
online job boards or social media platforms for a more extensive talent pool. Furthermore, the
system could evolve to provide predictive analytics for workforce planning, helping
organizations anticipate staffing needs and optimize resource allocation. Ongoing
advancements in facial recognition technology could be leveraged to enhance the accuracy and
efficiency of the attendance tracking system. Overall, the continuous evolution of this project
could position it as a cutting-edge solution for holistic talent management and organizational
efficiency.

Dept.of CSE-CBIT-Kolar 2 2023-24

HRMS WITH AI INTRODUCTION

1.4 OBJECTIVES
• The objective of this project is to develop a Flask web application that utilizes the
Resumeparse library to analyse resumes submitted by job applicants.
• The application will extract skills from both resumes and job descriptions, compare
them, and automatically shortlist resumes based on the compatibility of skills
Additionally, the system will incorporate face recognition technology to take
attendance of employees, providing a seamless and efficient method for tracking
workforce presence.
• The overarching goal is to streamline the recruitment process by automating the initial
screening of resumes while also enhancing attendance tracking through advanced
biometric technology.

1.5 ORGANIZATION OF THE PROJECT

The report is organized into chapters as follows:
Chapter 1- Introduction: This chapter presents a brief description about human resource
management with ai.

Chapter 2- Review of literature: In this section, the works carried out by various researchers.

Chapter 3- System requirement specification: The chapter 3 presents the specific

requirement, software and hardware requirements interfaces used. It also presents a brief
summary about the chapter.

Chapter 4- System analysis: The chapter 4 presents the Existing system and Proposed system.

BIBILOGRAPHY: 10 References.

APPENDIX A: Acronyms used in the report.

APPENDIX B: 10 Reference paper

Dept.of CSE-CBIT-Kolar 3 2023-24

CHAPTER 2
LITERATURE SURVEY
The literature survey serves as a critical foundation for this research, providing
a comprehensive overview of existing scholarly work and insights related to the specific
domain of study. By reviewing a diverse range of academic articles, books, and relevant
publications, the literature survey aims to identify key themes, methodologies, and gaps in the
existing knowledge base. This exploration not only contextualizes the current research within
the broader academic landscape but also facilitates the identification of areas where the
proposed study can contribute new perspectives or address unresolved questions. The literature
survey, therefore, plays a pivotal role in shaping the research framework, refining research
questions, and establishing the theoretical groundwork upon which the subsequent
investigation builds.

2.1 RELATED WORK

[1] The basis of our research is to construct a job recommendation system to the job seekers by
collecting the job portals data. Due to huge amounts of the data in job portals the employers
are facing difficulty in the identification of right candidate for the required skill and experience.
The job seekers are also facing the problem of getting the suitability of the job based on their
skill and experience. The knowledge acquisition based on the requirements is very difficult in
case of huge amounts of the data sources. In fact, classical development of domain ontology is
typically entirely based on strong human participation. It does not adequately fit new
applications requirements, because they need a more dynamic ontology and the possibility to
manage a considerable quantity of concepts that human cannot achieve alone. The main focus
of our work is to generate a job recommendation system with the details of job by taking
account into the data posted in the web sites and data from the job seekers by the creation of
dynamic ontology. We strongly believe that our system will give the best outcome in case of
suitable job recommendation for both employers and job seekers without spending much time.
To achieve this first we have extracted the data from various web pages and stored the collected
data into .csv files. In the second stage the stored input files are used by the similarity measure
and ontology creation module by generating the corresponding Web Ontology Language (.owl)
file. The third stage is creating the ontology with the generated. owl by using protégé tool.

[2] Haze poses a great challenge in modern-day applications. Removing haze is a challenging
task because haze varies with the depth of the scenes in the image. Many automated systems

4
HRMS WITH AI LITERATURE SURVEY

like surveillance systems, object tracking, etc. use the dehazing methods internally to improve
their overall performance in a hazy environment. Hence, recovering the original image from
the hazy image is a crucial task. This paper provides a detailed survey of many algorithms that
have been proposed for haze removal, which refines the image, either by using statistical
observation of the scene or by network-based learning method. These algorithms have been
assessed quantitatively and visually to compare based on their dehazing potential. The paper
presents a comparative study of the various image dehazing methods having a different way of
estimation of transmission and atmospheric light. This paper reviews different prior-based
technique and learning-based technique image dehazing algorithms.

[3] Job Recommendation system is a technique, which provides users with information,
which he/she may be interested in or accessed in past. Traditional recommender techniques
such as content and collaborative filtering used in various applications such as education, social
media, marketing, entertainment, e-governance and many more. Content-based and
collaborative filtering has many advantages and disadvantage and they are useful in specific
application. Sparsity and cold start problem are major challenges in content and collaborative
filtering. Challenges of content and collaborative filtering can be solved by using hybrid
filtering. Hybrid filtering combines the features of two recommender system like content
and collaborative; content-based filtering improves the classification accuracy and
collaborative model easily gives the best-predicted result of a latent factor model. In this
paper, we have presented a brief survey of the recommendation system approaches,
techniques and application, one important application of recommendation system in Job
Recruitment; in which candidates are elected by using online job recruitment portal based on
their profile and job history and behaviour components; wherein it serves millions of
candidates with suitable and personifies jobs. As per the recent survey this domain is less
explored till now and existing job recommender system has many shortcomings, they use
resumes/profile and job descriptions for analysis and new job post and candidate profiles
are not matched properly because of cold start problem, sometime potential candidate loses
their job due to the incomplete job description and education detail in the ontology.
LinkedIn’s Job Ecosystem handles few problems; few are still unsolved that we discussed
in result part.

[4] Job recommender systems are designed to suggest a ranked list of jobs that could be
associated with employee’s interest. Most of existing systems use only one approach to make
recommendation for all employees, while a specific method normally is good enough for a

Dept.of CSE-CBIT-Kolar 5 2023-24

HRMS WITH AI LITERATURE SURVEY

group of employees. Therefore, this study proposes an adaptive solution to make job
recommendation for different groups of user. The proposed methods are based on employee
clustering. Firstly, we group employees into different clusters. Then, we select a suitable
method for each user cluster based on empirical evaluation. The proposed methods include CB-
Plus, CF-jFilter and HyR-jFilter have applied for different three clusters. Empirical results
show that our proposed methods is outperformed than traditional methods.

[5] Dealing with the enormous amount of recruiting information on the Internet, a job seeker
always spends hours to find useful ones. To reduce this laborious work, we design and
implement a recommendation system for online jobhunting. In this paper, we contrast user-
based and item-based collaborative filtering algorithm to choose a better performed one. We
also take background information including students’ resumes and details of recruiting
information into consideration, bring weights of co-apply users (the users who had applied the
candidate jobs) and weights of student usedliked jobs into their commendation algorithm. At
last, the model we proposed is verified through experiments study which is using actual data.
The recommended results can achieve higher score of precision and recall, and they are more
relevant with users’ preferences before.

[6] In the strategic human resource (HR) management literature, over the past three decades, a
shared consensus has developed that the focus should be on HR systems rather than individual
HR practices because the effects of HR practices are likely to depend on the other practices
within the system. Despite this agreement, the extent to which the fundamental assumption in
the field of interactions and synergy in the system holds true is unclear. We present a systematic
review of 495 empirical studies on 516 HR systems in which we analyze the development of
HR systems research over time and identify important trends, explicitly linking
conceptualization and measurement of the HR system. Our findings suggest that the
increasingly broad conceptualization and measurement of HR systems and the lack of clarity
on the HR systems construct at different levels have hampered research progress. Much of the
research to date does not align with the fundamental assumption of synergies between HR
practices in a system, the measures have problems and increasingly confound HR systems with
related concepts and outcomes, and insufficient attention is paid to the HR system construct at
different levels. Overall, we thus still know little about the “systems” element and how
synergies and interactions in an HR system operate. We offer actionable suggestions on how
to advance HR systems research towards conceptual clarity and construct refinement, focusing

Dept.of CSE-CBIT-Kolar 6 2023-24

HRMS WITH AI LITERATURE SURVEY

both on how to conceptualize, measure, and combine practices in systems and on studying such
systems at different levels of analysis.
[7] The purpose is to strengthen Human Resources Management (HRM) through information
management using Artificial Intelligence (AI) technology. First, the selection criteria of the
applicant’s resume during recruitment and the formulation standards of the contract salary are
analyzed. Then, the resume information is extracted and converted into the data-type format.
Besides, the salary forecast model in the HRM system (HRMS) is designed based on the Back
Propagation Neural Network (BPNN), and network structure, parameter initialization, and
activation function of the BPNN are selected and optimized. The experimental results
demonstrate that the algorithm optimized by the Nadm has shown improved convergence speed
and forecast effect, with 187 iterations. Moreover, compared with other regression algorithms,
the designed algorithm achieves the best test scores. The above results can provide references
for designing the AI-based HRMS
[8] Artificial Intelligence (AI) is increasingly present in organizations. In the specific case of
Human Resource Management (HRM), AI has become increasingly relevant in recent years.
This article aims to perform a bibliometric analysis of the scientific literature that addresses in
a connected way the application and impact of AI in the field of HRM. The scientific databases
consulted were Web of Science and Scopus, yielding an initial number of 156 articles, of which
73 were selected for subsequent analysis. The information was processed using the
Bibliometrix tool, which provided information on annual production, analysis of journals,
authors, documents, keywords, etc. The results obtained show that AI applied to HRM is a
developing field of study with constant growth and a positive future vision, although it should
also be noted that it has a very specific character as a result of the fact that most of the research
is focused on the application of AI in recruitment and selection actions, leaving aside other
sub-areas with a great potential for application
[9] The study designs and implements the model of human position matching using ML
algorithms, which can provide a guarantee for effective human resource allocation. A new
method is proposed to evaluate the candidates in an online recruitment system and to solve the
problem of candidate ranking using ML algorithms. Further, a new algorithm based on support
vector regression is designed for the prediction of the human resource demand of enterprises
which has a good practical reference value. Based on the work and life situation, this study
analyzes the factors affecting the retirement of employees and uses ML algorithms to create a
retirement prediction model, which can provide solutions for the outflow of human resources.

Dept.of CSE-CBIT-Kolar 7 2023-24

HRMS WITH AI LITERATURE SURVEY

is study designs and implements the model of human position matching using ML algorithms,
which can provide a guarantee for effective human resource allocation. A new method is
proposed to evaluate the candidates in an online recruitment system and to solve the problem
of candidate ranking using ML algorithms. Further, a new algorithm based on support vector
regression is designed for the prediction of the human resource demand of enterprises which
has a good practical reference value.
[10] Digital innovation continues to fuel business transformation. Organizations have realigned
their strategic direction on enhanced adoption of digital technologies to leverage the
opportunities provided by the new age technologies, especially Artificial Intelligence.
Enterprises can leverage strategic advantage in talent-a key differentiator-by adopting Artificial
Intelligence in Human Resource Management. More than ever now, Human Resources
function now regarded as a trusted advisor, helping the organizations get through the
transformational phase created by disruptive technologies. This research provides insights on
how Human Resources function has evolved as a strategic partner by deployment of AI related
technological advancements related, as it contributes to building organizational capabilities and
making organization competitive, thus creating organizations that win in the market. It also
looks at the challenges faced in Human Resource Management by deployment of Artificial
Intelligence. Insights are shared on future directions of potential research that can be conducted
in this field.

Dept.of CSE-CBIT-Kolar 8 2023-24

CHAPTER 3

SYSTEM REQUIREMENT SPECIFICATION

System requirement specifications gathered by extracting the appropriate
information to implement the system. It is the elaborative conditions which the system need to
attain. Moreover, the SRS delivers a complete knowledge of the system to understand what
this project is going to achieve without any constraints on how to achieve this goal. This SRS
not providing the information to outside characters but it hides the plan and gives little
implementation details.

3.1 HARDWARE REQUIREMENTS

1. RAM: Minimum of 2 GB
2. Processor: AMD Ryzen 7 5700U with Radeon Graphics 1.80 GHz or Any Processor
with the speed more than 2.5GHz

3.2 SOFTWARE REQUIREMENTS

1. Operating System: Windows 8 and Later Version: The web application is designed
to run seamlessly on Windows 8 and later versions, ensuring compatibility with a
widely used operating system. This choice allows organizations and users with
Windows-based environments to easily deploy and integrate the application into their
existing infrastructure. Windows compatibility also ensures a familiar and user-friendly
experience for administrators and other users interacting with the system.

2. Backend Language: Python with SQLite3: The backend of the web application is
developed using Python as the primary programming language. Python is known for its
readability, versatility, and extensive library support, making it an excellent choice for
web development. The SQLite3 database is employed to store and manage data
efficiently. SQLite is a lightweight and serverless database engine, making it suitable
for smaller-scale applications and projects. It simplifies the deployment process and
contributes to the portability of the application.

3. Frontend Languages: HTML, CSS, JavaScript: The frontend of the web application
is built using a combination of HTML, CSS, and JavaScript. HTML provides the
structure of the web pages, CSS handles the styling and layout, and JavaScript adds
interactivity and dynamic functionality. This triad of frontend technologies ensures a
responsive and visually appealing user interface, offering a smooth and engaging

9
HRMS WITH AI SYSTEM REQUIREMENTS

experience for recruiters, administrators, and other users interacting with the
application through their web browsers.

4. Framework: Python Flask: The Python Flask framework is chosen to build the web
application's backend. Flask is a lightweight and modular web framework that
simplifies the development of web applications in Python. Its simplicity, flexibility, and
ease of integration make it well-suited for projects of varying sizes. Flask follows the
WSGI standard and supports extensions, allowing developers to customize and scale
the application as needed. The use of Flask streamlines the development process and
contributes to the overall efficiency of the web application.

5. Tools: Visual Studio, Python IDE: The development environment for the web
application involves the use of Visual Studio and a Python Integrated Development
Environment (IDE). Visual Studio provides a comprehensive integrated development
environment with features like code editing, debugging, and version control. The
Python IDE enhances the development experience specifically for Python code,
offering tools and features tailored to the language. The combination of these tools
accelerates the development cycle, aids collaboration among developers, and ensures
the creation of a robust and well-maintained web application.

3.3 FUNCTIONAL REQUIREMENTS

1. User Authentication:
• The system should have secure user authentication for administrators,
recruiters, and employees.
• Different user roles (admin, recruiter, employee) with specific privileges and
access levels.
2. Resume Parsing:
• Utilize the Resumeparse library to automatically analyse resumes submitted by
candidates.
• Extract and store relevant information such as skills, qualifications, work
experience, and education.
3. Job Posting:
• Allow recruiters/administrators to create and post job openings, specifying job
requirements and qualifications

Dept.of CSE-CBIT-Kolar 10 2023-24

HRMS WITH AI SYSTEM REQUIREMENTS

• Link job requirements to specific skills extracted from resumes.

4. Resume-Requirement Matching:
• Implement a matching algorithm to compare skills extracted from resumes with
the job requirements.
• Automatically shortlist resumes that closely match the specified job criteria.
5. Candidate Management:
• Enable recruiters/administrators to view and manage candidate profiles.
• Provide features for adding notes, comments, and feedback on each candidate.
6. Face Recognition for Attendance:
• Integrate a face recognition system to automate employee attendance tracking.
• Store attendance records for each employee securely.

3.4 NON-FUNCTIONAL REQUIREMENTS

1. Security:
• Implement robust security measures to protect sensitive data, including user
authentication, data encryption, and secure connections.
2. Scalability:
• Design the application to handle a growing number of users, resumes, and job
postings without significant performance degradation.
3. Usability:
• Create an intuitive and user-friendly interface for recruiters and administrators.
• Provide clear instructions and guidance on using the platform.
4. Reliability:
• Ensure high system availability and minimize downtime.
• Implement regular backups to prevent data loss.
5. Performance:
• Optimize the application for fast response times, especially during resume
parsing and matching processes.
6. Compatibility:
• Ensure compatibility with different web browsers and devices to accommodate
a diverse user base.

Dept.of CSE-CBIT-Kolar 11 2023-24

CHAPTER 4
SYSTEM ANALYSIS
System analysis is the process of studying a procedure or business to identify
its goal and purposes and create systems and procedures that will efficiently achieve them.
Another view sees system analysis as a problem-solving technique that breaks down a system
into its component pieces, and how well those parts work and interact to accomplish their
purpose.

4.1 EXISTING SYSTEM

In existing project, the studies of different types of recommendation systems are
presented. The methods include collaborative filtering, content-based filtering, personalized
and non-personalized algorithms the advantages and disadvantages described. The
comparative study of various techniques mentioned above is presented in this report. The data
set used for the projects, the implementation. As a result, students are able to make an informed
choice and are walked through the course which is well suited for them, this helps the students
to polish their skills and perform better in academics. It also helps with placements and other
competitive examinations.

4.2 PROPOSED SYSTEM

The proposed system is a Flask-based web application designed for efficient
recruitment and employee management. Leveraging the Resumeparse library, the system
enables users to upload resumes, automatically extracting relevant skills and information for
analysis. The application then compares these extracted skills with the specified job
requirements, assisting in the shortlisting process. Additionally, the system incorporates face
recognition technology to facilitate employee attendance tracking, ensuring a streamlined and
secure approach to workforce management. Through this integrated solution, the proposed
system aims to enhance the recruitment process by automating resume analysis and simplify
employee attendance tracking through facial recognition, ultimately improving overall
organizational efficiency.

12
APPENDIX A

ACRONYMS

1. WSGL - Web Server Gateway Interface

2. CSS - Cascading Style Sheet
3. HTML - Hypertext Markup Language
4. HRMS - Human Resource Management System
5. AI - Artificial Intelligence
6. CF - Collaborative Filtering

13
 BIBILOGRAPHY
[1] Kethavarapu U. P. K., Saraswathi S. Concept based dynamic ontology creation for job
recommendation system. Procedia Computer Science . 2016;85:915–921. doi:
10.1016/j.procs.2016.05.282. [CrossRef] [Google Scholar]
[2] Parihar A. S., Gupta Y. K., Singodia Y., Singh V., Singh K. A comparative study of
image dehazing algorithms. Proceedings of the 2020, 5th International Conference on
Communication and Electronics Systems (ICCES); 2020, June; Coimbatore, India. pp.
766–771. [CrossRef] [Google Scholar]
[3] Mishra R., Rathi S. ICDSMLA 2019 . Singapore: Springer; 2020. Efficient and scalable
job recommender system using collaborative filtering; pp. 842–856. [Google Scholar]
[4] Mhamdi D., Moulouki R., El Ghoumari M. Y., Azzouazi M., Moussaid L. Job
recommendation based on job profile clustering and job seeker behavior. Procedia
Computer Science . 2020;175:695–699. doi: 10.1016/j.procs.2020.07.102. [CrossRef]
[Google Scholar]
[5] Desai V., Bahl D., Vibhandik S., Fatma I. Implementation of an automated job
recommendation system based on candidate profiles. Int. Res. J. Eng. Technol .
2017;4(5):1018–1021. [Google Scholar]
[6] P.R. Palos-Sánchez, P. Baena-Luna, A. Badicu & J.C. Infante-Moro (2022)Artificial
Intelligence and Human Resources Management: A Bibliometric Analysis,
AppliedArtificial Intelligence, 36:1, 2145631, DOI: 10.1080/08839514.2022.2145631
[7] A Systematic Review of Human Resource Management Systems and Their
Measurement Corine Boon,Deanne N. Den Hartog ,University of Amsterdam ,David
P. Lepak University of Massachusetts Amherst.
[8] Artificial Intelligence in Human Resource Management - Challenges and Future
Research Recommendations Article · (July 2023) DOI: 10.24818/RMCI.2023.3.382.
[9] Research on Human Resource Recommendation Algorithm Based on Machine
Learning, Hong Zhu 1,2 ,Xi’an Jiaotong University, Xi’an, Shaanxi 710049, China
,Shaanxi XueQian Normal University, Xi’an, Shaanxi 710100, China Correspondence
should be addressed to Hong Zhu; [email protected] .Received 21 June 2021;
published 25 August 2021,Academic Editor: Muhammad Usman
[10] Design and interactive performance of human resource management system based on
artificial intelligence Yangda Gong , Published in year:2022

14
APPENDIX B

REFERENCE PAPERS

15
 Available online at www.sciencedirect.com

ScienceDirect
Procedia Computer Science 175 (2020) 695–699

The International Workshop on Artificial Intelligence and Smart City Applications

(IWAISCA) August 9-12, 2020, Leuven, Belgium

Job Recommendation based on Job Profile Clustering and Job

Seeker Behavior
D. Mhamdi*, R. Moulouki, M. Y. El Ghoumari, M. Azzouazi, L. Moussaid
Mathematics, Computing and Information Processing, Faculty of Sciences Ben M'Sik, Hassan II University, Casablanca 20670, Morocco

Abstract

This article presents a recommender system that aims to help job seekers to find suitable jobs.
First, job offers are collected from job search websites then they are prepared to extract
meaningful attributes such as job titles and technical skills. Job offers with common features
are grouped into clusters. As job seeker like one job belonging to a cluster, he will probably
find other jobs in that cluster that he will like as well. A list of top n recommendations is
suggested after matching data from job clusters and job seeker behavior, which consists on
user interactions such as applications, likes and rating.

© 2020 The Authors. Published by Elsevier B.V.

This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the Conference Program Chair.
Keywords: job recommendation; profile clustering; word2vec; k-means clustering.

1. Introduction
In the era of Big Data, both of employers and job seekers are confronted with the increasing
data overload and the time-consuming process of recruitment. Candidate’s profiles are so
diverse that it is laborious for recruiter to find the suitable competencies. Consequently, it is
important to identify the most important features of each job candidature.
Job recommendation system are machine learning solutions capable of suggesting pertinent
jobs or candidates based on the behavior and needs of job seekers and on requirements of
employers. Therefore, applicants could receive personalized online jobs and recruiters are
supposed to find the most relevant candidatures with skills and qualifications that fit their
needs.

16
1877-0509 © 2020 The Authors. Published by Elsevier B.V.

This is an open access article under the CC BY-NC-ND license

(http://creativecommons.org/licenses/by-nc-nd/4.0/) Peer-review under responsibility of the
Conference Program Chairs.
10.1016/j.procs.2020.07.102
696

Fig. 1. Content based filtering and collaborative filtering recommendation [3].

Natural language processing as a branch at the intersection of computer science, artificial

intelligence, and linguistics could be used to extract useful insights from job offers in order to
match candidates to suitable offers. In addition, it offers the capabilities of processing and
analyzing large quantities of unstructured job postings and job applications as to provide
recruiters and job applicants the ability to understand the preferences and requirements of each
other. As a result, they can save time and efforts.
In this paper, first, we have presented related works concerning automated recommendation
and some text clustering methods, and then we have exposed the basis and rules of our
proposed model.

2. Related works
2.1. Automated Recommendation

While conducting a search on the web, users are supported by automated recommendation to
find and choose the right items that fit their needs, according to people they trust or sharing
similar tastes. Automated recommendation is divided into content-based filtering and
collaborative filtering [1].
As shown Fig. 1, Content-based methods suggest items similar to those a user has selected in
the past. Collaborative filtering recommend objects based on the preferences of other users
with tastes similar to those of the current user [2].
Word2vec: a neural network based model considering words from a corpus and representing
them as vectors with contextual comprehension. Two words are considered similar if the
distance between their respectively vectors is lower [5].

17
 Word2vec is a model used to produce word embedding, which is an embedded representation
of documents that consists on mapping words or phrases from a vocabulary to corresponding
vectors of real numbers. Words that appear together in the text will also be very close in vector
space [6].

697

Fig. 2. Difference between SkipGram and CBOW [7].

As shown in Fig. 2, Word2vec includes two architectures of performing distributed

presentation of words: continuous bag-of-words (CBOW) and skip-gram.
While the CBOW architecture predicts the current word based on the context, the
Skip-gram predicts surrounding words given the current word.

• K-means clustering: K-means is a popular and simple to implement algorithm. It groups N

data points into k clusters by minimizing the sum of squared distances between every point
and its nearest cluster mean called centroid. It starts by selecting k random data points as
the initial set of centroids of clusters. Then the centroid of every cluster is recalculated as
the mean of all data points assigned to this cluster [8].
There is a variant of k-means algorithm called Spherical k-means considering all vectors as
normalized and using cosine dissimilarity based on the angle between the vectors as calculated
in formulae (1) presented below [9]:

< 𝑥𝑥, 𝑝𝑝 >

d(x, p) = 1 − cos(x, p) = 1 − (1)
||𝑥𝑥|| ||𝑝𝑝||

18
The spherical K-means has as input a set of unit vectors that lie on the surface of the unit
hypersphere about the origin, and uses the cosine similarity as its proximity measure [10].

3. Proposed recommender system

Our Dataset includes job offers and job seeker interactions such as rating, likes and reviews.
The objective is matching job offers to the right candidates.
A job posting is a document made up of structured data such as the position title and
unstructured data such as the map of a location. Many features describe a job offer including
the company name, the job title, the job description, the location, the post time, the salary, and
the job requirements.
The job title and the job description are main attributes that hold meaningful data. The title is
a short and succinct text that designates the job position. The description of the job contains
all the skills, qualifications and conditions related to the job.

698

Fig. 3. Matching job offers to job seekers.

The data that will be processed and analyzed is collected from different job search websites
using web-scraping techniques.
As shown in Fig. 3, the output of our model is a list of sorted recommendations. Many steps
are to perform including data collection, textual processing and matching profiles. In a first
step, the unstructured data from job offers is transformed and validated to make it more easily
understood. Then, the formatted and cleaned data is processed and analyzed in order to be
divided into job clusters based on common features. The attributes from the data contained in
these clusters are matched with behavior attributes of the job seekers and a list of n
recommendations is suggested to the user. When a job offer is liked or rated by a candidate,
all relevant job offers belonging to the same cluster are suggested to the same candidate.
This model is based on cluster analysis approach, which is a self-organized learning that helps
to identify groups of job offers according to the degree of similarity, or dissimilarity between
their features.

19
 4. Conclusion and future work
In this paper, we presented a job recommender model aiming to extract meaningful data from job
postings using text-clustering methods. As a result, job offers are divided into job clusters based on
their common features and job offers are matched to job seekers according to their interactions.
Our future Work will focus on training and evaluating our model using Word2vec method and k-
means clustering algorithms used to capture and represent the context of job profiles. Subsequently,
it will be easy to match set of job offers to a given job seeker based on its past interactions toward
specific job offers. The dataset that will be used is built from scraping job search websites.

References
[1] Frank Faeber, Tim Weitzel, and Tobias Keim. (2003) “An Automated Recommendation
Approach to Selection in Personnel Recruitment.” Americas Conference on Information
Systems (AMCIS).
[2] Marko Balabanovic, and Yoav Shoham. (1997) “Content-based, Collaborative
Recommendation.” Communications of the Association for Computing Machinery 40(3),
66-72.
[3] Marwa Hussien Mohamed, Mohamed Helmy Khafagy, and Mohamed Hassan Ibrahim.
(2019) “Recommender Systems Challenges and Solutions Survey.” International
Conference on Innovative Trends in Computer Engineering, Aswan, Egypt.
[4] Hoajun SUN, Zhihui LIU, and Lingjun KONG. (2008) “A Document Clustering Method
based on Hierarchical Algorithm with Model Clustering.” 22nd International Conference
on Advanced Information Networking and Applications.
[5] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. (2013) “Efficient
Estimation of Word Representations in Vector Space.” Proceedings of the International
Conference on Learning Representations.
[6] D. Mhamdi, R. Moulouki, M. Y. El Ghoumari, and M. Azzouazi. (2020) “Job
Recommendation System based on Text Analysis.” Jour of Adv Research in Dynamical &
Control Systems, Vol. 12, 04-Special Issue.
[7] Kalyan Katikapalli Subramanyam, and Sangeetha Sivanesan. (2019) “SECNLP: A Survey
of Embeddings in Clinical Natural Language Processing.” Journal of Biomedical
Informatics.
[8] Pasi Franti, and Sami Sieranoja. (2019) “How much can k-means be improved by using
better initialization and repeats?” Pattern Recognition 93 (2019) 95-12, Machine Learning
Group, School of Computing, University of Eastern Finland.
[9] Kurt Hornik, Ingo Feinerer, Martin Kober, and Christian Buchta. (2012) “Spherical k-
Means Clustering.” Journal of Statistical Software, Vol. 50, Issue 10.
[10] Rehab Duwairi, and Mohammed Abu-Rahmeh. (2015) “A novel approach for initializing
the spherical k-means clustering algorithm.” Simulation Modelling Practice and Theory
54 (2015) 49-63.

20
 A Comparative Study of Image Dehazing
Algorithms
Anil Singh Parihar, Yash Kumar Gupta, Yash Singodia, Vibhu Singh, Kavinder Singh
Machine Learning Research Lab, Dept. of Computer Science
Delhi Technological University
Delhi, India
[email protected], [email protected], [email protected],
[email protected],
[email protected]
Abstract—Haze poses a great challenge in [6]–[14], and Low-light enhancement [15]–
modern-day applications. Removing haze is [17] are a part of Image Enhancement.
a challenging task because haze varies with
the depth of the scenes in the image. Many The visibility of outdoor regions is
automated systems like surveillance deteriorated in adverse weather conditions
systems, object tracking, etc. use the with the presence of haze, fog, water
dehazing methods internally to improve droplets and various different particles in
their overall performance in a hazy the atmosphere. The presence of haze is a
environment. Hence, recovering the
dominant issue that decreases the contrast
original image from the hazy image is a
crucial task. This paper provides a detailed level and brings the noise to the image.
survey of many algorithms that have been Consequently, haze removal and improving
proposed for haze removal, which refines the colour of the image becomes a difficult
the image, either by using statistical task. The volume of the haze or fog depends
observation of the scene or by network- on the floating particles in the environment.
based learning method. These algorithms These bits of particle absorb and scatter
have been assessed quantitatively and light, which reflects back from scene
visually to compare based on their dehazing
objects. The degree to which light scatters,
potential. The paper presents a comparative
study of the various image dehazing rely on the separation gap between the
methods having a different way of object and the observer. Object tracking
estimation of transmission and atmospheric [18], Object detection [19], Action
light. This paper reviews different prior- recognition [20], surveillance systems
based technique and learning-based ,environment monitoring and other
technique image dehazing algorithms. automated systems require clear and haze-
Index Terms—Image dehazing, learning- free images to work properly. Moreover, It
based, CNN, Enhancement, Atmospheric
is difficult to perform any other vision-
scattering model
based task effectively in a hazy image.
Thus, Haze removal becomes a necessity.
.
Many researchers [21], [22] performed haze
I. INTRODUCTION removal with the help of multiple images.
Image enhancement endeavours to improve the These methods attempt to take more than
visual quality of an image. Image can degrade one scene-image with an intent to learn the
due to lighting, atmosphere, device issues and different features of the image. Separate
many other reasons. Researchers presented images contain different features and thus,
different approaches to deal with different can help in capturing haze relevant feature
issues and achieve image enhancement. effectively. However, it is sometimes not
Contrast enhancement [1]– [5], Haze removal possible to have multiple images for feature

21
extraction. It is a challenging task to distil camera without getting dispersed and L tells
out pertinent haze features from a single the atmospheric light. Also, the
input image because of very less transmission is proportionate to the
information provided by a single image. exponential distance between object and
Various researchers [7], [8], [23]–[25] lens, therefore
endeavoured to single image haze removal.
T = e−βD (2)
The dehazed outputs produced by many
single image dehazing algorithms appeared here β is the atmosphere coefficient of
pleasing to the eye and the overall scattering and D denotes the distance
complexity also got reduced. Hence, between the camera and the object. The
addressing this problem is important in model is widely employed in image
computer vision to recover images from the dehazing to produce enhanced results R
haze regions for useful information. The from input I by evaluating T and L.
paper presents a contrasting examination of
several image dehazing algorithms. III. LITERARY SURVEY
The main objective of image dehazing
Section-II presents Atmospheric scattering
methods is to produce outputs that contains
model. SectionIII contains Brief discussion
appropriate visible object in the image with
of various single image dehazing methods.
almost no noise. This paper presents a brief
Section-IV propounds a detailed
deliberation of various image dehazing
comparative analysis. Section-V is used for
methods that are commonly used. Image
the conclusion.
dehazing can be used in variety of
II. ATMOSPHERIC SCATTERING applications. Various methods are proposed
MODEL by researchers to remove haze from an
image and generate images with better
The Atmospheric scattering model [6] is
image quality. The analysis contains prior-
extensively exploited to deal with hazy
based and learning-based approaches. In the
images. The model is stated as:
following section, various algorithms will
Ic = RcT + Lc(1 −T) (1) be discussed.

A. DARK CHANNEL PRIOR

here I depicts input-image, R denotes the
The dark channel prior (DCP) [7] forms on
scene radiance of the image, c denotes
the statistical knowledge of outside scene
colour channels, and Transmission T
without fog. It is built on the inspection that
depicts quantity of light that falls straight to
the local patchs Ω, which are not a part of
lens of the
sky region, have some pixels, with less
intensity values in no less than one of the
colour channel. These low intensities are
accounted by the presence of shadows,
colourful objects and dark surfaces. DCP is
built on atmospheric model of image
formation (1). The approach estimates
Atmospheric light using dark-channel
which is given as:
Fig. 1. Atmospheric Scattering Model [26]
Rdark = min (minRc) (3)

22
 cε(R,G,B) yεΩ DCP generates artifacts while dealing with
images having white object. In DCP,
where Rdark denotes dark-channel, and Ω transmission is dependent on atmospheric
denotes a patch size. After estimation of light and prior. Thus, a improved prior can
dark-channel, it estimates Transmission provide better enhancement.
map as:
T = 1 − ω min(min(Ic/Lc)) (4) B. COLOR ATTENUATION PRIOR
Colour attenuation prior (CAP) [8] is built
c yεΩ on the scrutiny that, the difference in the
intensity of brightness and that of an image
saturation varies hazy regions of an image
where T denotes Transmission map, and Lc
than hazefree regions greatly. For hazy
denotes the channel-wise atmospheric light. parts, this difference is vast, whereas, the
DCP, assumes that the Rdark of a haze free difference is low for the haze-free regions.
Hence, It could be inferred that D∝ ρ ∝B -
image is nearly 0, and Lc is always non- S, where D symbolize the depth of scene, ρ
negative. is the haze concentration in the image, S is
the intensity of saturation whereas B is the
Rdark → min (minRc(y)) (5) intensity brightness of the scene. CAP
estimates depth map using the following
cεR,G,B yεΩ learning model:
Since, the Complete elimination of haze D(p) = δ0 + δ1B(p) + δ2S(p) + ε(p)
may appear displeasing to the eye. Thus, (8)
DCP preserves few traces of haze in the
p is the position, in the image, δ0,δ1,δ2 are
image to take care of the aerial perspective
unknown parameter and θ0 is equal to 1.0,
for distant objects.
θ1 is equal to -1.0 and is a input image. The
It is done by introducing a new parameter depth map can be estimated using equation
ω, which is set to (8). CAP estimates of atmospheric light by
choosing the brightest 0.1% pixel in the
0.9.
depth and selecting the pixels intensity
T = 1 − ω min(min(Ic(y)/Lc)) corresponding to them in the input image.
Transmission map can be estimates by
(6) c yεΩ
using Eq (3). Hence, The scene radiance can
The scene radiance can be recovered by be estimated as:
using estimated Transmission, and L
Atmospheric light. Hence, It computes the R − (9)
final dehazed image as:
It limits transmission map between a lower
L R(7) bound and upper bound to achieve natural
where t0 denots the lower bound for enhancement. CAP performs better than
Transmission, which keeps transmission at DCP as it uses depth for transmission
particular lower limit. estimation. However, it sometimes provides
DCP work appropriately when images do limited enhancement. Since, it uses hard
not contain any white or brighter object. thresholding to limit the scene radiance.

23
 C. RANKING CNN Ranking CNN attempts to learn haze
Ranking Convolutional Neural Networks characteristics using ranking layer. The
[11] attempts to cover the drawbacks of the other techniques are unable to distinguish
results attained from [9] and [10].Any statistical and structural details in the
image has two features that play a major image. Ranking CNN evaluates statistical
role in the formation of transmission map. details with structural details from an
They are called the structural and the image. While the other methods were taking
statistical features. Some of the finest CNN many traversals of the filters to calculate the
may attain success in evaluating the statistical details, the ordering of elements
structural features but most of them in the layer aids in the calculation of
statistical details like contrast easily and
does not require many traversals. This
approach does not assumes that the
transmission is constant locally instead it
aims to find transmission at every pixel of
the image, thereby increasing the overall
image enhancement ability.

Fig. 2. Calculation of statistical feature D. DEHAZE NET

(contrast) [11] DehazeNet [10] was proposed which being
fall short while estimating statistical details. a [CNN] could be trained for memorizing
The Ranking layer was proposed keeping haze characteristics and the effects they
both these features in mind as it can have in transmission. DehazeNet attempts
estimate both the above said features. to learn associations between hazy areas
Ranking Layer takes feature maps as input and their transmission. The attributes such
and then return an output of ranked maps as Dark Channel, Maximum Contrast,
with the same shapes and sizes. The role of Colour Attenuation and Hue Disparity, if
this layer is just to change the ordering of removed from an image makes the image
the input map. This aids the overall haze-free. To improve the chances of
dehazing process by increasing its convergence this method also brings with it
efficiency in calculation of features. a novel non-linear activation function
known as Bilateral-Rectified Linear Unit
It can be observed from the 2 that for the (BReLU).
calculation of the statistical feature other
CNN may need many convolution filters The structure of DehazeNet is a
but the ranking layer needs only one combination of many convolutional,
filter.This Ranking Layer is then inserted pooling layers stacked back to back along
into the classical CNN to get an estimation
over the pixel-wise transmission. The
remaining part is a regression network
followed by a Random Forest generator to
increase performance. The obtained result
is then passed for smoothing and rectified
result is then coalesced with L, which is
evaluated by calculating the 0.1% pixels Fig. 3. Architecture of DehazeNet [10]
having the highest intensities in the Dark with the presence of activation functions as
Channel, to produce final results. shown in Fig. 3 The main feature of

24
DehazeNet includes haze relevant acquired by the forward and backward
characteristic extraction from the whole iterations of the network. Hence, it provides
image using a non linear regression based a better haze representation as compared to
on BReLU. Network learns the training other dehazing techniques.
parameters through Stochastic Gradient
Descent (SGD) and Mean-Squared Error F. AOD NET
(MSE) is its loss function. The end result of All-In-One Dehazing Network (AOD-Net)
this network is the efficient calculation of [12] evaluates haze-free results directly
transmission map. from the foggy image without separate
evaluation of transmission and airlight.
During the estimation of transmission near
AOD Net changes the atmospheric model
the sky region, some of the values of the
into a K-module based model. The
outliers in CAP [8] fluctuates. However,
DehazeNet is able to calculate the
transmission with greater precision. Also,
CAP performance degrades badly when the
concentration of haze increase which is not
observed in DehazeNet. Further, BReLU is
an efficient activation function that helps in
image restoration and conversion.

E. MULTISCALE CNN
To evaluate the transmission MSCNN [9]
Fig. 4. MSCNN model [9]
presented a two step approach. First, a
rough outline depicting the transmission is
obtained utilizing a Coarse-Scaled
Network. The produced rough result
containing the overall details of the image
is then passed along to the second phase.
Convolution, pooling, upsampling are some
of the operations performed during the first
Fig. 5. Architecture of AODNet [12]
phase. Second, the Fine-Scaled Network
enhances the overall structure of the
intermediate passed onto it. The coarse K-module based model incorporates the
transmission is provided here as an add-on effect of the transmission, and airlight,
feature map. The transmission here gets Using Eq. (1).
refined due to the features learned by the
fine-scaled model through forward and R (10)
backward iterations. MSCNN model is as The purpose here is to entangle L and into
shown in Fig. 4. K.
Some methods depends on prior based R = KI−K + b (11)
knowledge and fails as compared to
MSCNN because MSCNN tries to learn the where b has default value 1 and is a
haze relevant features with the help of an bias.Here,
end to end CNN. MSCNN method includes
the effects of priors and other details K (12)

25
 I− 1 represents the architecture of the model
which contains 4 modules : 1. Pyramid
After the model is converted as per the K Structure Transmission Estimation Map. 2.
module equation, the next step is the clean Atmospheric light estimation module. 3.
image conversion module that incorporates Overall Dehazing. 4. GAN Joint
many pixel wise additions and Discriminator .
multiplication layers. The results then
It optimizes GAN using the edge-
obtained as the dehazed outputs. Because of
preserving loss function. In practice they
the fact that the dehazed output is produced
concatenated images free from fog with the
by a single network run, AOD Net has its
transmission and then leveraged use of
application that it can be inserted with other
joint-discriminator loss function as:
learning based models for faster overall
results. Fig. 5 shows AOD Net framework. Gt,Gd Dj oint ∼pdata (I) [log(1 − Dj
oint(Gt(I)))]+ min max EI
AOD Net is robust in cases when there are
objects with colour alike atmospheric light.
In those cases DehazeNet [10], MSCNN [9]
show very less details. Also, the artifact Several CNN-based methods like Ranking-
problems are present in DCP [7], CAP [8], CNN [11], MSCNN [9], restricts their
DEHAZENet [10], MSCNN [9], however ability by only taking into account of
AOD Net is artifact free method. transmission map for their CNN framework
whereas AODNet [12] solves this issue
G. DCPDN (Densly Connected Pyramid with their light-weight CNN.
Dehazing Network)
Zhang, and Patel [13] deveoped an end-to- H. FFA-NET
end model for removal of haze using Qin et al. [14] presented a fusion end-to-end
Generative Adversarial Network(GAN). attention model for image dehazing. It uses
The model learns structural relation from Channel Attention and Pixel Attention
the images. The model consists of encoder- modules which considers features based on
decoder structure which can learn different channels. The use of residual
learning is stated to bring less vital features
like thin haze. Feature Fusion Module is
used to learn different features on different
channels. It provides more weight to
important features like thick haze regions
with fusion of weights from different
channels.
Fig. 7 shows Feature Attention module in
Fig. 6. Architecture of DCPDN Model [13] FFA-Net. It consists of with basic residual
transmission, airlight and can dehaze the learning which consists N group
image simultaneously. Transmission is architectures which learns the less
better calculated with the help of multipool important features like low frequency and
mapping. Also, atmospheric model has hazy region to avoid through multiple skip
been included for better optimization for connections. The feature attention module
overall learning process. DCPDN has less improves the performance of the algorithm
ability for being temporally coherent. Fig. 6 which consists of channel and pixel

26
attention modules. Channel attention relies as thin haze and thick haze should have
on the fact that the channels of image different weighting factors and with the
consists of different features and heterogeneous nature of haze distribution
spatialinformation is obtained by using along different channels are also taken in
global average pooling, with consideration.
IV Quantitative Analysus

This section presents the analysis of above-

discussed methods. For computing
differences and evaluate efficiencies among
different dehazing methods, both
quantitative and qualitative analysis are
necessary. Fig. 8 depicts the outputs that are
produced by distinct dehazing techniques.
Fig. 8(a) shows the input image with haze.
convolution layers and element wise
Fig. 8(b-h) represents output of He [7], Zhu
multiply with the weights of channels as:
[8], Li [12], Ren [9], Cai [10], Qin [14],
Zhang [13] respectively. It can be noted that
Fig. 8(b) is removing haze efficiently but at
times suffer from halo-artifacts near the
boundaries of the subject such as tree or
F *c = CAc ⊗ F c (16) building. Fig. 8(c) does not suffer from
excess saturation and brings out the real
,where Ic denotes cth channel of Image with colours. The outputs of Fig. 8(d) have
pixel position at (i,j), Hp indicates the global sharper details and maintains the colour
average pooling layer, σ is a sigmoid accuracy of the image.In the presence of
function, δ denotes the ReLu function, CAc objects having colour close to white light,
denotes the channel weights which is this method shows robustness and performs
multiplied with input Fc. efficiently. Fig. 8(f) eradicates haze nicely
and the colour distribution is also uniform
Pixel Attention module is further used by
thereby producing good results. But as the
the network to learn more information
number of hazy images increases, the visual
about thick hazy regions. It consists of
evaluation becomes harder and biased.
convolution lanes and element-wise
Therefore it is required to present the
multiplication and outputs the Feature
quantitative evaluation of the above-
Attention module.
mentioned methods. Learning-based
PA = σ(Conv(δ(Conv(F *)))) techniques require a lot of data for learning
(17)
complex features so after research, the
F˜ = F * ⊗ PA (18)
datasets that are used in these methods for
,where PAc denotes output of Pixel learning haze relevant features are:
Attention, which is further element-wise RESIDE [29], NYU Depth v2 [30],
multiplied to obtain Feature Attention . Middlebury Stereo [31], NTIRE [32],
FRIDA [33], IHAZE [34], and O-HAZE
Previous CNN-based models like Ranking-
[35]. The codes and weights given by
CNN [11], DehazeNet [10], MSCNN [9],
authors are used for fair comparison.
AOD-Net [12] handle the attention modules
Further, the image quality is tested using
for channel and pixel-wise as uniform, but,

27
metrics such as Structure Similarity [2] A. S. Parihar, “Histogram modification
Index(SSIM) [27], Peek Signal-to-Noise and dct based contrast enhancement,” in
Ratio(PSNR), Weighted Peek-Signal-to- 2018 2nd International Conference on
Noise Ratio(WPSNR) [28] , and Mean Inventive Systems and Control (ICISC),
Squared Error. Since these are widely used 2018, pp. 614–618.
metrices for analyzing image dehazing [3] A. S. Parihar, “Fuzzy adaptive gamma
methods. The synthetic image data is used correction for contrast enhancement,”
from the project sites of [36], [37] to gather in 2017 International Conference on
Intelligent Sustainable Systems (ICISS),
the result with application of those metrics.
2017, pp. 625–629.
Quantitative results are compared in Table
[4] A. S. Parihar, O. P. Verma, and C.
1. The value of these metrics are averaged
Khanna, “Fuzzy-contextual contrast
out for the images of data set. enhancement,” IEEE Transactions on
V. CONCLUSION Image Processing, vol. 26, no. 4, pp.
1810–1819, 2017.
After analyzing Image dehazing
[5] A. S. Parihar, O. P. Verma, and D.
approaches, it can be observed that each
Yadav, “Image contrast enhancement
approach contains their own limitations.
using differential evolution,” in
This paper depicts an extensive inspection Advances in Communication, Devices
of various state of art image dehazing and Networking. Singapore: Springer
approaches. It is observed that the statistical Singapore, 2018, pp. 517–526.
methods fail in regions where the object’s [6] E. J. McCartney, “Optics of the
colour was identical to the atmospheric atmosphere: scattering by molecules
light. Statistical methods performs better in and particles,” New York, John Wiley
quantitative analysis. However, in and Sons, Inc., 1976. 421 p., 1976.
Qualitative analysis these methods shows [7] K. He, J. Sun, and X. Tang, “Single
various artifacts. The algorithms either image haze removal using dark channel
produced a halo effect or erased the object. prior,” IEEE Transactions on pattern
The learning based methods performed analysis and machine intelligence, vol.
efficiently on most of the images. The 33, no. 12, pp. 2341–2353, 2010.
contrast of learning based methods is [8] Q. Zhu, J. Mai, and L. Shao, “A fast
similar to the statistical methods. The single image haze removal algorithm
learning based approach performs using color attenuation prior,” IEEE
appreciably in qualitative analysis. The Transactions on image processing, vol.
24, no. 11, pp. 3522–3533, 2015.
performance of statistical methods can be
[9] W. Ren, S. Liu, H. Zhang, J. Pan, X.
improved using improved prior knowledge
Cao, and M.-H. Yang, “Single image
and parameters. Learning-based methods
dehazing via multi-scale convolutional
can overfit on a dataset thus, it must neural networks,” in European
consider the overfitting conditions conference on CV. Springer, 2016, pp.
appropriately. 154–169.
REFERENCES [10] B. Cai, X. Xu, K. Jia, C. Qing, and D.
Tao, “Dehazenet: An end-to-end
[1] A. S. Parihar and O. P. Verma, system for single image haze removal,”
“Contrast enhancement using IEEE Transactions on Image
entropybased dynamic sub-histogram Processing, vol. 25, no. 11, pp. 5187–
equalisation,” IET Image Processing, 5198, 2016.
vol. 10, pp. 799–808(9), November [11] Y. Song, J. Li, X. Wang, and X. Chen,
2016. “Single image dehazing using ranking

28
[12] convolutional neural network,” IEEE [21] S. Ji, W. Xu, M. Yang, and K. Yu, “3d
Transactions on Multimedia, vol. 20, convolutional neural networks for
no. 6, pp. 1548–1560, 2017. human action recognition,” IEEE
[13] B. Li, X. Peng, Z. Wang, J. Xu, and D. transactions on pattern analysis and
Feng, “Aod-net: All-in-one dehazing machine intelligence, vol. 35, no. 1, pp.
network,” in IEEE Conference on 221–231, 2012.
Computer Vision, 2017, pp. 4770–4778. [22] Y. Y. Schechner, S. G. Narasimhan, and
[14] H. Zhang and V. M. Patel, “Densely S. K. Nayar, “Instant dehazing of
connected pyramid dehazing network,” images using polarization,” in IEEE
in IEEE Conference on CVPR, 2018, Conference on CVPR, vol. 1. IEEE,
pp. 3194–3203. 2001, pp. I–I.
[15] X. Qin, Z. Wang, Y. Bai, X. Xie, and H. [23] Y. Y. Schechner, S. G. Narasimhan, and
Jia, “Ffa-net: Feature fusion attention S. Nayar, “Polarization-based vision
network for single image dehazing,” through haze,” Applied optics, vol. 42,
arXiv preprint arXiv:1911.07559, no. 3, pp. 511–525, 2003.
2019. [24] J.-P. Tarel and N. Hautiere, “Fast
[16] K. Singh and A. S. Parihar, “A visibility restoration from a single color
comparative analysis of illumination or gray level image,” in IEEE
estimation based image enhancement Conference on Computer Vision, 2009,
techniques,” in 2020 International pp. 2201–2208.
Conference on Emerging Trends in [25] G. Meng, Y. Wang, J. Duan, S. Xiang,
Information Technology and and C. Pan, “Efficient image dehazing
Engineering (ic-ETITE), 2020, pp. 1–5. with boundary constraint and
[17] A. S. Parihar and K. Singh, “A study on contextual regularization,” in IEEE
retinex based method for image conference on computer vision, 2013,
enhancement,” in 2018 2nd pp. 617–624.
International Conference on Inventive [26] C. O. Ancuti and C. Ancuti, “Single
Systems and Control (ICISC), Jan 2018, image dehazing by multi-scale fusion,”
pp. 619–624. IEEE Transactions on Image
[18] A. S. Parihar and K. Singh, Processing, vol. 22, no. 8, pp. 3271–
“Illumination Estimation for Nature 3282, 2013.
Preserving low-light image [27] C.-C. Tsai, C.-Y. Lin, and J.-I. Guo,
enhancement,” “Dark channel prior based video
10.36227/techrxiv.12236780.v1, 5 dehazing algorithm with sky
2020. [Online]. Available: preservation and its embedded system
https://www.techrxiv.org/articles/Illum realization for adas applications,”
ination Estimation for Nature Optics express, vol. 27, no. 9, pp.
Preserving low-light image 11877–11901, 2019.
enhancement/12236780 [28] Z. Wang, A. C. Bovik, H. R. Sheikh,
[19] L. Bertinetto, J. Valmadre, J. F. and E. P. Simoncelli, “Image quality
Henriques, A. Vedaldi, and P. H. Torr, assessment: from error visibility to
“Fully-convolutional siamese networks structural similarity,” IEEE
for object tracking,” pp. 850– 865, transactions on image processing, vol.
2016. 13, no. 4, pp. 600–612, 2004.
[20] J. Redmon, S. Divvala, R. Girshick, and [29] J. Mannos and D. Sakrison, “The
A. Farhadi, “You only look once: effects of a visual fidelity criterion of
Unified, real-time object detection,” pp. the encoding of images,” IEEE
779–788, 2016. transactions on Information Theory,
vol. 20, no. 4, pp. 525–536, 1974.

29
[30] B. Li, W. Ren, D. Fu, D. Tao, D. Feng,
W. Zeng, and Z. Wang, “Benchmarking
single-image dehazing and beyond,”
IEEE Transactions on Image
Processing, vol. 28, no. 1, pp. 492–505,
2018.
[31] N. Silberman, D. Hoiem, P. Kohli, and
R. Fergus, “Indoor segmentation and
support inference from rgbd images,” in
European conference on CV. Springer,
2012, pp. 746–760.
[32] H. Hirschmuller and D. Scharstein,
“Evaluation of cost functions for stereo
matching,” in IEEE Conference on
CVPR. IEEE, 2007, pp. 1–8.
[33] C. Ancuti, C. O. Ancuti, and R.
Timofte, “Ntire 2018 challenge on
image dehazing: Methods and results,”
in IEEE Conference on CVPR
Workshops, 2018, pp. 891–901.
[34] J.-P. Tarel, N. Hautiere, L. Caraffa, A.
Cord, H. Halmaoui, and D. Gruyer,
“Vision enhancement in homogeneous
and heterogeneous fog,” IEEE
Intelligent Transportation Systems
Magazine, vol. 4, no. 2, pp. 6–20, 2012.
[35] C. O. Ancuti, C. Ancuti, R. Timofte,
and C. D. Vleeschouwer, “I-haze:
a dehazing benchmark with real hazy
and haze-free indoor images.”
[36] C. O. Ancuti, C. Ancuti, R. Timofte,
and C. De Vleeschouwer, “O-haze: a
dehazing benchmark with real hazy and
haze-free outdoor images,” in IEEE
CVPR, 2018, pp. 754–762.
[37] R. Fattal, “Dehazing using color-lines,”
ACM transactions on graphics (TOG),
vol. 34, no. 1, pp. 1–14, 2014.
[38] D. Berman, S. Avidan et al., “Non-
local image dehazing,” in Proceedings
of the IEEE conference on CVPR, 2016,
pp. 1674–1682.

30
 Available online at www.sciencedirect.com ScienceDirect

Procedia Computer Science 85 ( 2016 ) 915 –

921

International Conference on Computational Modeling and Security (CMS 2016)

Concept Based Dynamic Ontology Creation for Job

Recommendation System
Uma Pavan Kumar Kethavarapu a Dr.S.Saraswathib
a
Research Scholar,CSE Department,Ponicherry Engineering
College.,Pondicherry-605001, India bProfessor and HOD IT
Department, Pondicherry Engineering College, Pondicherry,-
605001, India

Abstract
The basis of our research is to construct a job recommendation system to the job seekers by
collecting the job portals data. Due to huge amounts of the data in job portals the employers
are facing difficulty in the identification of right candidate for the required skill and
experience. The job seekers are also facing the problem of getting the suitability of the job
based on their skill and experience. The knowledge acquisition based on the requirements
is very difficult in case of huge amounts of the data sources. In fact classical development
of domain ontology is typically entirely based on strong human participation. It does not
adequately fit new applications requirements, because they need a more dynamic ontology
and the possibility to manage a considerable quantity of concepts that human cannot
achieve alone. The main focus of our work is to generate a job recommendation system
with the details of job by taking account into the data posted in the web sites and data from
the job seekers by the creation of dynamic ontology. We strongly believe that our system
will give the best outcome in case of suitable job recommendation for both employers and
job seekers without spending much time. To achieve this first we have extracted the data
from various web pages and stored the collected data into .csv files. In the second stage the
stored input files are used by the similarity measure and ontology creation module by
generating the corresponding Web Ontology Language (.owl) file. The third stage is
creating the ontology with the generated .owl by using protégé tool .

© 2015 The Authors. Published by Elsevier B.V.

© 2016 The Authors. Published by Elsevier B.V. This is an open access article under the
CC BY-NC-ND license
Peer-review under responsibility of organizing committee of the 2016 International
Conference on Computational Modeling and (http://creativecommons.org/licenses/by-nc-
nd/4.0/).
Peer-review under responsibility of the Organizing Committee of CMS 2016Security (CMS 2016).

Keywords:Ontology;Job Portals;Job Seekers;Web Ontology Language;ElasticSerach

31
1877-0509 © 2016 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer-review under responsibility of the Organizing Committee of CMS
2016 doi: 10.1016/j.procs.2016.05.282

Introduction
Ontology is a formal explicit specification of shared conceptualization which will provide
Common semantics for agent communication. Currently, applications mostly exchange
information on the basis of passing parameters or data, formatted according to pre-defined
strict syntaxes this approach is known as the exactness method 1. This method has the
advantage of allowing total error management, except application bugs of course, but leaves
no space for data interpretation. In consequence, reasoning on data of this type is virtually
impossible because of the limits of its definition. Ontologies provide a richer knowledge
representation that improves machine interpretation of data. For this they become to be widely
used in information systems and applications, and ontology construction has been addressed
in several research activities. Ontology must be able to grow dynamically without deviation
of the existing applications. At the same time computational time for discovering the best
matches between several ontologies is expensive, therefore the technique must maintain
previous discovered alignments and common usages in order to quickly recognize similarities
between concepts and to compute only new information. Ontology is designed not only to
provide a complete view of domain concepts but also to identify quickly and accurately
similarities between concepts, even if not identical, and to conduct consistent alignments 2.
The method we are working on gives the answer for the statement that is within the specified
domain without taking much time will give the best matching of the job recommendations by
considering the previous details and similarities of the key skills or work locations of the job
seekers.
Method Extraction Analysis Generation Validation Evolution
Generating an Human - Direct transformation Human -
ontology from an using XSLT files.
annotated business
model
XML2OWL Static table of - Direct transformation Human
correspondences using XSLT files.
UML2OWL Using Semi-automatic - Direct transformation Human
approach using XSLT files.
TERMINAE NLP techniques Concept No standard ontology Human -
relationships representation
analysis
SALT NLP techniques. Multi Similarity analysis of No standard ontology –Limited human -
entries. concepts representation intervention
A new Method for - Based on semi- Automatic merging. -
Ontology Merging automatic ontology No standard ontology
based on Concept representation.
using WordNet
Design of the Main concept defined - Human Intervention -
Automatic by a domain expert.
Ontology Building
System about the
Specific Domain
Knowledge
Enriching Very Enrich existing ontology - Human Intervention -
Large Ontologies
Using the WWW

32
Ontologies can be built as manually, semi-automatically or automatically. If ontologies are
constructed manually then we can term that ontology as manual ontology, while constructing
the ontologies if human intervention is required then we can name that as semi-automatic
ontologies, if the system takes care about complete construction of ontologies then we can say
that as automatic ontology. A collection of documents that are related to the user’s request are
produced by comparing the user’s request with an automatically generated index of the textual
content of the documents present in the system by means of a computerized process, called
document Retrieval (also known as Information Retrieval) Text retrieval in which information
is basically stored in the form of text is a branch of information retrieval.
In some cases there may be a chance of changing the requirements in the input to the system.
In such cases we must use the dynamic ontologies so as to embed the changes in the existing
result sets. The organization of the paper is as follows in the section 2 we are going to explain
the methods available in automatic ontology construction. In section 3 the details about various
automatic ontology comparisons, in section 4 the explanation about the job recommendation
system that we are constructed in section 5 future scope and conclusion is explained.

2 Tools Of Automatic Ontology Construction

In the construction of Automatic ontology five parameters are playing a vital role, those are
Extraction, Analysis, Generation, Validation and Evolution. The following table gives
the analysis of various tools based on the above mentioned parameters.

Proposed Research
Based on the above analysis we are going to construct one automatic ontology with similarity
measure threshold value and re ranking method. The process we have adopted for the
construction of automatic ontology for the job recommendation system is as follows. Initially,
the web log data can be collected from the Job portals. After the raw data can be collected, the
pre- processing can be done. Initially, ontology creation and mapping will be done by
analysing various properties of ontologies in order to deduce alternate semantics that may
apply to other ontologies, and therefore create a mapping. The feature extraction module based
on TF-IDF similarity, and then Indexing and ranking of information by Rabin Fingerprint
algorithm and ranking can be done by semantic similarity measure 13. Once the ontology is
mapped, the relevant information will be retrieved effectively based on the user query. That is
for the input query keyword, matching will be performed with the mapped ontology and the
exact information’s will be retrieved using the computed similarity score. Lastly, by means of
the matching result the related report were generated from the document repository. The
implementation is done in Java with protégé 14, 15 software tool for ontology creation and
updating. The performance of the proposed system will be analysed in terms of accuracy, recall
and F-measure. Extraction: The source of extracting the data is web pages and we are using
the domain of job portals such as naukri, monster and times job. The procedure here is we will
input URL name the web page and the Read Data procedure will generates the .csv file. The
same procedure will be repeated to multiple web pages so as to generate the .csv files. All the
generated files will be collected in an input directory. Analysis: In our system the analysis part
is pre-processing such as identification of the URL information based on the requirement like
skill, experience and qualification, actually this step is the basis for the generation of the .csv
files forms the web portals. Generation: The generation of the .csv files and after that by using
re-ranking and similarity measures we are generating the .owl file which is the source file to

33
the ontology construction, after populating the .csv files based on those files the Run() method
integrates the multiple .csv files and filters the required records with similarity measure and
re-ranking approaches. Once this process is done then the Run () will generates the newtest.owl
file. Based on the .owl file we can automatically construct the ontology with the specified
attributes of the job recommendation system.
Validation: The validation process is done manually by the user after generation of the
ontology by using protégé tool.
Evolution: The evolution process is by default property in our system reason is that as we are
using the streaming data in the form of web portals the information is up to date and dynamic
so based on that only the generation of .csv and .owl is done, with the generated .owl we are
generating the ontology with protégé tool.
Accuracy: In addition to the above parameters we are considering the one more parameter that
is accuracy of the result. In our implementation we are using similarity measure along with
threshold value by comparing the word net source of the data as a dictionary for searching the
words, will give some value based on that we can generate the records as per our requirements.
The following functionalities we have used in the implementation
FindMeasure (Query, data, originaldata)-Gives the measure value of the given data by
comparing with the original data.
SelectBest (data [], score, originaldata)-Gives the best score for the given data in the query
from the original data. Score Value: The score value is the deciding factor for the generation
of relevant records and this will works based on the user given threshold value like score>0.1
Some of the mathematical computations performed in our implementation of work. tf (t in d)
correlates to the term's frequency, defined as the number of times term t appears in the
currently scored document d 19.

(1)

idf(t) stands for Inverse Document Frequency. This value correlates to the inverse of docFreq
(the number of
documents in which the term t appears). This means rarer terms give higher contribution
to the total score. idf (t) appears for t in both the query and the document; hence it is squared
in the equation 20.

(2) In the measuring of the relevancy of the records populated

by the .csv files we are using score value of the records generated in such a way that , a
threshold value we will mention to get more relevant records with similarity measure. Here
the measure will compute the score value out of the existing records and their attributes with
the help of word net repository 14, 15 and it outputs the most relevant records with the help of
that only the .owl will be generated.
The data set considered for the research is web portal data such as naukri, monster and times
jobs etc., from these portals we are generating the .csv and the system configuration is HP with
8GB RAM and Core I3 processor with Net Beans 8.1.The other soft wares required to test.The
further research of our work requires Python, MongoDB along with elasticsearch so as to get
the global job recommendation systems. The term frequency in our research is used to get the

34
number of times a term occurred in a document. For example “Java” is a term which is
occurred 158 times, ” Python” is term which occurred 185 times and the term “Hadoop”
occurred 152 times, machine learning occurred 3 times. We have used the terms like “Python”
“Bangalore” ,”Java”, ”Chennai” and “Hadoop” ,”Mumbai” together so as to get the resultant
records. Another computing we have done is inverse document frequency where we can get
the score value of the term in a document. For example if I am giving java then the idf value
is 0.5235 and similarly for Python I am getting the value of 0.85732, where as in case of
machine learning it is like 0.03010.Based on the score value we are going to compare with
minimum threshold so as to get the more relevancy of the records. The threshold value we
have taken in the experiment is 0.01 and we can see the outcome of the query based on the
input query along with the threshold value. The following is the outcome of the similarity
measure in the implementation of proposed research. Once the generation of .owl is over the
next step is to generate the ontology with set of attributes and the relationships. Initially the
entire data items are converted into ontology format and there after we can query the ontology
based on the requirements. In the following diagram we can see the attributes and relationships
surrounded by ontology and the export of ontograph is possible for the query where workplace
is equal to “Bangalore” .The following is the sample diagram of OntoGraph generated for the
query where work place=”Bangalore”.

Fig. 1. Ontology Graph With .OWL using Protégé Tool.

The following is the outcome of the similarity measure in the implementation of proposed
research. Once the generation of .owl is over the next step is to generate the ontology with
set of attributes and the relationships. Initially the entire data items are converted into
ontology format and there after we can query the ontology based on the requirements.

35
 Fig. 2.Similarity Measure computing based on the given query Python
Bangalore.

The following is exported version of the DOT file from the created ontology.
digraph g {
"trendwiseanalytics" -> "Bengaluru_Bengaluru Bangalore" [label="in_place"]
"place" -> "Bengaluru_Bengaluru Bangalore" [label="has individual"]
"jobname" -> "Urgent_opening_jr Java_Developer_bangalore" [label="has
individual"]
"Arnold_Consulting_Private_Limited" -> "Urgent_opening_jr
Java_Developer_bangalore" [label="has_vacancy"]
}

Future Scope and Conclusion

The overall theme of our work is to construct the automatic ontology with in less time and
more accurate. The domain we are working is job recommendation system through which we
are taking the input as web pages and generation of the .csv files. In the next stage these
multiple .csv’s are used to generate the .owl file. And in the generation of ontology we are
giving some query so as to output the records. At this stage we are using similarity measure
based on the term frequency and inverse document frequency which is helpful in the
generation of content with more relevant data. The generated ontology is competent on par
with manual ontology in the context of accuracy and similarity of the data generation. And the
benefits of our work compared with manual ontology construction is time saving and complete
automation of tasks like .csv generation , getting the .owl from multiple
.csv’s and finally generation of automatic ontology. The elastic search 21 and integration other
social media sites information to the job seekers data is our next focus .The expected outcome
is to filter the irrelevant and fake resumes if any conflicts or any negative results we are getting.
The methodology we are going to adopt is exhaustive search 22 on the social sites along with
the job portals so that one global job recommendation system of notifications generation for
both the employers and job seekers.

References

[1] Nelia Lasierra,”Designing an Architecture for Monitoring Patients at Home:

Ontologies and Web services for Clinical and Technical
[2] Management Integration,” IEEE JOURNAL OF BIOMEDICAL AND HEALTH
INFORMATICS, VOL. 18, NO. 3, MAY 2014 p 10-24. 2. Mark Roantree,
Automating the Integration of Clinical Studies into Medical Ontologies”, 2014 47th
Hawaii International Conference on System Science p 190-205.
[3] Vishal Jain, “Architecture Model for Communication between Multi Agent Systems
with Ontology”, International Journal of Advanced Research in Computer Science,
Volume 4, No. 8, May-June 2013 p 123-134.

36
[4] Pavel A. Smirnov”, Domain Ontologies Integration for Virtual Modelling and
Simulation Environments”, Procedia Computer Science, Elsevier, Volume 29, 2014,
Pages 2507–2514.
[5] S.MONISHA,” A FRAMEWORK FOR ONTOLOGY BASED LINK
ANALYSISFOR WEB MINING”, Journal of Theoretical and Applied
[6] Information Technology, 20 thMarch 2015. Vol.73, No.2 p 32-45.
[7] Vishal Jain,” MINING IN ONTOLOGY WITH MULTI AGENT SYSTEM IN
SEMANTIC WEB: A NOVEL APPROACH”, The
[8] International Journal of Multimedia &Its applications (IJMA) Vol.6, No.5, October
2014 45-55.
[9] Fernández López, M.,” Overview Of Methodologies For Building Ontologies”,
Proceedings of the IJCAI-99 workshop on Ontologies and Problem-Solving Methods
(KRR5)Stockholm, Sweden, August 2, 2013.
[10] Pratibha Gokhale,” Ontology Development Methods”, DESIDOC Journal of
Library & Information Technology, Vol. 31, No. 2, March 2011, pp. 77-83.
[11] Ying Ding”, Ontology Research and Development – A Review of Ontology
Mapping and Evolving”, Journal of Information Science, 28(5), 375-388, 2012.
[12] Helena Sofia Pinto”, A Methodology for Ontology Integration”, K-CAP’01,
October 22-23, 2001, Victoria, British Columbia, Canada, Copyright 2001 ACM 1-
58113-380-4.
[13] Trong Hai Duong,” Complexity Analysis of Ontology Integration
methodologies: a Comparative Study”, Journal of Universal Computer Science, vol.
15, no. 4 (2009) p 18-25
[14] Mark Roantree”, Automating the Integration of Clinical Studies into Medical
Ontologies”, IEEE Computer Society, 2014.
[15] YuxiaHuang”, Using Ontologies and formal concept analysis to integrate
heterogeneous tourism information “, IEEE Transactions in Emerging Topics in
Computing, June 2015 p 25-33.
[16] O. Iroju,” State-of-the Art: A Comparative Analysis of Ontology Matching
Systems “, African Journal of Computing &ICT, Vol 5. No. 4, June 2012 ,IEEE ISSN
2006-1781.

37
 Efficient and Scalable Job Recommender
System Using Collaborative Filtering
Ravita Mishra (&) and Sheetal Rathi
Thakur College of Engineering and Technology, Kandivali (E) Mumbai 400101,
India
[email protected], [email protected]

Abstract. Recommendation system is a technique, which provides users with information,

which he/she may be interested in or accessed in past. Traditional recommender techniques
such as content and collaborative filtering used in various applications such as education,
social media, marketing, entertainment, egovernance and many more. Content-based and
collaborative filtering has many advantages and disadvantage and they are useful in specific
application. Sparsity and cold start problem are major challenges in content and collaborative
filtering. Challenges of content and collaborative filtering can be solved by using hybrid
filtering. Hybrid filtering combines the features of two recommender system like content and
collaborative; content-based filtering improves the classification accuracy and collaborative
model easily gives the best-predicted result of a latent factor model. In this paper, we have
presented a brief survey of the recommendation system approaches, techniques and
application, one important application of recommendation system in Job Recruitment; in
which candidates are elected by using online job recruitment portal based on their profile and
job history and behaviour components; wherein it serves millions of candidates with suitable
and personifies jobs. As per the recent survey this domain is less explored till now and existing
job recommender system has many shortcomings, they use resumes/profile and job
descriptions for analysis and new job post and candidate profiles are not matched properly
because of cold start problem, sometime potential candidate loses their job due to the
incomplete job description and education detail in the ontology. LinkedIn’s Job Ecosystem
handles few problems; few are still unsolved that we discussed in result part. In this paper, we
have presented a comparative analysis of different job recommender system and their
techniques.

1 Introduction
Now-a-day social media is very common platform to share the data and day today’s activities.
With the enormous use of various internet sources likes, mobile phone and smart devices,
Internet users can receive huge information about shopping and social activity of user and
online learning.
If the data volume and variety increases tremendously, then individual user faces various
problem of excessive information, it causes problem to make the correct decisions. This
framework is called as information overload. To resolve users’ information overload problem,
a new technique recommender system comes in pictures.Recommender system can solve
various problems by effectively finding users’ probable requirements and elect fascinating
items from a vast amount of applicant information. Recommender systems are mainly
categorized into three main forms,i.e., content-based (CB), collaborative filtering (CF) and

38
hybrid recommender system is combination of both resolve the drawback of content and
collaborative filtering

requirements and p to determine whether the item is useful for the user. Demographic systems
(DG) recommends items on the basis of user’s demographic profile, it helps in marketing to

Recommender
system

Content based Collaborative Hybrid

Filtering Filtering Filtering

Model Memory
based based

Classification
Clustering

recommend different items [1] . Fig. 1 depicts the different approach of recommender system
.

Fig. 1. Approaches of recommender system

In computer network application, social network generates a large volume of data on a daily
basis and social networks, because of large size and structure the integrate growing field of
research.
This worth of data is leveraged by any recommender systems, and resulting findings can help
to solve the interesting problems related to social obligation, recruitment, and in friend
suggestion. One of important application of recommender system is job recommendation it
recruits suitable candidates through job portal. Every day, lots of candidates browse online job
portal for finding the effective, meaningful and trans-parent job. Now-a-days many online job
portals are available and they used content based or collaborative recommendation. One of job
recommender is work4 which is San Francisco based company and uses the concept of content
based and offers Facebook recruitment solution by taking some attribute of then. This system
leads many challenges like sparsity of job and user, overspecialization and limited content
analysis. The existing also not solved the unmatched jobs/candidates pair, many low
qualifications of applicants that match the search criteria [2]. The best fit between jobs and
candidate pairs are solely depends on underlying aspects that are hard to measure. To solve the
existing recommendation system’s problem by merging the important feature of content and
collaborative filtering and they solve individual’s problem each other’s. In hybridization we
used the features of both approaches and dropping their limitation, it gives better result as
compare to earlier approach. Content based system referred many algorithms like K nearest
neighbour, clustering, Bayesian clustering and deep learning. In collaborative filtering uses user
based and item based algorithm such as support vector machine, principal component analysis,
and neural networks. Hybrid system uses all existing algorithm such as feature combination,
maximum entropy and deep learning.
This paper gives immense knowledge of accuracy measure, application domain; challenges of
existing job recommender system and comparison between different job recommender systems.
We have compared different methods, approach and evaluation matrix of a recommendation

39
system of three companies’ dataset and their results are discussed. The paper is organized in
five parts, the Sect. 1 contain introduction about recommender system, Sect. 2 gives the state
of arts, Sect. 3 gives various research questions and Sect. 4 gives the job recommendation
comparison, Sect. 5 presents the approaches, comparison of the different algorithm and finally
concludes this paper.

2 Literature Survey

A detailed survey on recommendation system is provided by Kumar and Sharma [3], Eirinaki
et al. [4], presented an immense idea about different taxonomy of recommender system, gaps
exists in current system, application domain such as television, research agency, restaurant
booking, job search and many more new research areas. Author also highlighted recommender
systems in large scale social media and summarizes different issues, unusual problems and their
solutions. Sun et al. [5], Yangi et al. [2] describe the video recommendation for online social
groups and has many limitations such as; group preferences and relationship between groups
members are not included in existing approaches. Author also address these issues, and presents
a R. Mishra and S. Rathi social aware group recommendation framework and use two important
characteristics of group members are tolerance and altruism are presented. Mobile internet
user’s problem can handle with the help of collaborative filtering (CF) and hybrid filtering and
also predict the interests of mobile users for effective recommendation purpose. Anandhan et
al. [6] presents different types of Recommendation system for online resources, such as blogs,
forums, social networking websites, bookmarking websites, and video and chat portals. Author
also presents recommendation system approaches, research domains; data sets used in each
domain, mining algorithm, recommendation type, and different performance measures. Wu et
al. [7] presents the real time mobile system Kaleido, uses a clustering approach and latent bias
model which gives recommendation solution by taking affective text into user account.
Adomavicius and Tuzhilin [8], Yang et al. [9] presents the state of arts of recommender system
and different techniques such as matrix factorization based and neural network system used for
online application such as voting system. Author also explores the social connection of user
and group affiliation information.
Aghasian et al. [10] presents the surveys on privacy concerns, measurements and privacy-
preserving techniques that will used in online social networks and extensive recommender
systems. Author discussed the privacy-preservation models for the users and provides
information sanitization and data obfuscation to assure data anonymity of individuals, address
the resource consumption and computation time and modeling attacks. Taghavi et al. [11],
Valverde-Rebaza et al. [12] presents the deep investigation of the real world recommender and
focus on the statistical research of published recommender systems, and also introduces the
taxonomies driven by several phases: initial phase, design and development phase, estimation,
function phase. Author pre-sents a model that used publicly available a new dataset formed by
a set of job seekers profiles and a set of job vacancies collected from different job search engine
sites; and put forward the proposal of a framework for job recommendation based on
professional skills of job seekers. Diaby et al. [13] presents another type of recommender
system (Job Recommender system) and offer job based on Facebook extracted field. With the
help of support vector machine and content based filtering analyses different field of resume
and Facebook are extracted.
Geyik et al. [14] Presents a LinkedIn Talent Solutions and provides a facility for job recruiter
to find out to suitable candidates and other hand for job aspirant to search best career

40
 opportunities. Author also addresses the traditional search and recommendation problem, for
best job opportunity they only focus on mutual interest between the recruiter and the candidate,
other problems are recruiter are not able to express their job hiring needs because of lack of
domain knowledge, time and manual effort for best search standard. Jalili et al. [15], Kenthapadi
et al. [16] presents, the statistical modeling system and few challenges of design and
implementation stage of the existing system, describes few modeling components (Bayesian
hierarchical and smoothing techniques) and its main task to presenting robust compensation to
users. Author also highlighted the different challenges and their solution like data collection
and pro-cessing; Improvement in statistical smoothing technology and improving outlier
detection at the initial stages.
LinkedIn Talent system addresses all these issues.

3 Research Questions

RQ1: What kinds of approaches are used for generating recommendations by Recomme nder
System?
The widely used recommendation technologies in commercial applications should be broadly
characterize as content based and rating based filtering.
Collaborative Recommendation (CF) technique gives recommendation based on items and
target user and consider past history of similar users with similar choices [15]. CF techniques
are further categorize into two main sub parts: memory-based that make the recommendations
compares a user’s historical records to other records in the database; and model-based uses
statistical learning, model fitting done by user directory then use it to generate predictions.
Content-Based Filtering (CB) gives recommendation based on items similar to the user has
liked in the past. Here profile is created for each individual user or item and describe its
important characteristics. For example, the attributes of a movie profile may be its genre,
director, story, actors, its box office popularity, effects etc. Hybrid Filtering (HF) follows a
blended approach that covers all other basic approaches to achieve some synergy among them.
Hybrid recommendation technique gives best performance by mitigating the
shortcoming/disadvantages of one technique using the strengths/advantages of other
techniques. In literature, there are many different ways proposed to create a hybrid system by
merging two or more basic approaches and several other approaches were also found in
literature that include data filtering techniques, used to extract information from data and can
be used to improve RSs performance. In 2002, Burke presented a many hybridization
techniques and they are classified as: Weighted, Switching, Mixed, Feature Segmentation, Meta
level and Cascade [ 17].
RQ2: What are the strengths and weaknesses of Recommendation System? Table 1 gives idea
about strength and weaknesses of algorithm and data in recommendation system and it is
indicated by possible (Y), not possible (N).
RQ3: Issues and Challenges encountered in Recommender System.
This section describes the most common issues and challenges that encounter in deploying RSs.

1. Sparse RSs: Generally, user does not rate product or new item which first time in market and
the resultant ratings matrix becomes very sparse. Due to this, sparsity
problem of data, it declines the chances of finding a set of users with similar
ratings.

41
2. Cold-start problem: It is also referred as new item and new user problem. In job
recommendation it is consider as new job and new candidate, recent jobs can’t be
recommended initially when it is introduced to a CF system with no ratings.
3. Scalability Problem: Scalability problem mainly arise in huge and dynamic data sets which is
produced by interactions between user and item such as preferences, ratings and reviews. It is
possible that when some recommendation algorithms are applied on relatively small data sets,
they provide the best results, but may reflect
inefficient or worst behaviour on very large datasets.
4. Privacy Issue: To produce quality personalized recommendations, RSs are bound to
gather as much user data as possible and to exploit it to the fullest.

42
 International Research Journal of Engineering and Technology (IRJET) e-ISSN:
2395 -0056
Volume: 04 Issue: 05 | May -2017 www.irjet.net
p-ISSN: 2395-0072

Implementation of an Automated Job Recommendation

System Based on Candidate Profiles
Vinay Desai, Dheeraj Bahl, Shreekumar Vibhandik,Isra Fatma
Prof. Vijay Kolekar , Dept. of Computer Engineering, K.J. College of Engg & Mgt. Research, Pune,
India
---------------------------------------------------------------------***------------------------------------------------------------
---------

Abstract - Dealing with the enormous online jobseekers are college graduates or
amount of recruiting information on the post graduates. The key problem is that
Internet, a job seeker always spends hours most of job-hunting websites just display
to find useful ones. To reduce this recruitment information to website viewers.
laborious work, we design and implement Websites just display recruitment
a recommendation system for online information to website viewers. Students
jobhunting. In this paper, we contrast user-
have to retrieve among all the information
based and item-based collaborative
to find jobs they want to apply. The whole
filtering algorithm to choose a better
performed one. We also take background procedure is tedious and inefficient. By
information including students’ resumes creating an easy job recommendation
and details of recruiting information into system where everyone will have a fair and
consideration, bring weights of co-apply square chance. This saves a lot of potential
users (the users who had applied the time and money both on the industrial as
candidate jobs) and weights of student well as the job seeker’s side. Moreover, as
usedliked jobs into their commendation the candidate gets a fair chance to prove his
algorithm. At last, the model we proposed talent in the real world it is a lot more
is verified through experiments study efficient system. The basic agenda of every
which is using actual data. The algorithm used in today’s world be it a
recommended results can achieve higher traditional algorithm or a hybrid algorithm
score of precision and recall, and they are
is to provide a suitable job that the user
more relevant with users’ preferences
before. actually seeks and wishes for.
1.1 LITERATURE SURVEY
Key Words- recommendation system;
item-based There are endless algorithms to help a
collaborative filtering; content-based seeker find the right job, some are the
filtering; Vector traditional algorithms while some are
Space Model(VSM); Mahout newly found and there are a large number
of hybrid algorithms which are a
1. INTRODUCTION combination of many algorithms. All these
algorithms have only goal to seek a
The increasing usage of Internet has righteous job for the candidate.
heightened the need for online job hunting.
According to Jobsite’s report 2014, 68% of

43
 International Research Journal of Engineering and Technology (IRJET) e-ISSN:
2395 -0056
Volume: 04 Issue: 05 | May -2017 www.irjet.net
p-ISSN: 2395-0072

CF is a popular recommendation algorithm quite simplified technique; all it does is

that bases its predictions and examine the past interests of the user and
recommendations on the ratings or based on the past results the system makes
behaviour of other users of the system. an accurate result of the candidate who has
applied for a job.
It also uses a profound technique called as
Information Retrieval (IR). Information Item-based CF: Rather than using
Retrieval is a new and advanced technique similarities between users’ rating behavior
used for achieving the most accurate and to predict preferences, item–based CF uses
desired result without compromising on the similarities
efficiency of the result.

• Between the rating patterns of items. Since

1.2 RELATED WORK finding similar items is easier than finding
similar users, and attributes of items are
A. Recommendation Algorithms more stable than users’ preference, item-
based methods are suitable for off-line
1) Content-based filtering (CBF): computing. The preferred outcome of any
In Content-based methods, features of items of the two methods is to provide a suitable
are abstract and compared with a profile of job. There are some drawbacks in
the user’s preference. In other words, this Collaborative Filtering as a whole aspect.
algorithm tries to recommend items that are Collaborative Filtering approaches often
similar to those that a user liked in the past. suffer from three problems: cold start,
It is widely applied in information scalability and sparsely. These three
retrieval(IR). However it performs badly in drawbacks are very problematic at times
multimedia field such as music or movie and crucial opportunities can be missed
recommendation because it is hard to extract because of this.
items attributes and obtains user’s
preference sometimes. B. Methods of Similarity Calculation 1)
Cosine Similarity
Cosine similarity uses two N-dimensional
2) Collaborative Filtering (CF): vector’s cosine Value to indicate the degree
CF is a popular recommendation algorithm of similarity between them. It is widely
that bases its predictions and used in information retrieval (IR).
recommendations on the ratings or behavior
of other users in the system. There are two
basic types: 2)Tanimoto Coefficient: Tanimoto
coefficient, also known as the Jaccard
User-based CF and Item-based CF. index,measures similarity between finite
• User-based CF: find other users whose sample sets, and is defined as the size of the
past rating behavior is similar to that of the
current user and use their ratings on other intersection divided by the size of the union
items to predict what the current user will of the sample sets.
like. The working of User –based CF is a
XY

44
International Research Journal of Engineering and Technology (IRJET) e-ISSN:
2395 -0056
Volume: 04 Issue: 05 | May -2017 www.irjet.net
p-ISSN: 2395-0072

Jaccard (X, Y) =X Y(2) items. Since finding similar items is easier

than finding similar users, and attributes of
3) Log Likelihood
items are more stable than users’
Similar to Tanimoto coefficient, the Log
preference, item-based methods are
likelihood method calculate similarity
suitable for off-line computing. The
based on the common preference two
preferred outcome of any of the two
users shared. Given the total number of
methods is to provide a suitable job. The
items and the number of each user rated
procedure is presented below:
items, the final result is the impossibility
of that the two users have such common for each jobi useri applied{ for
preference. each co-applied userj who
applied jobi{ find out jobs that
userj applied; add these jobs to
4) The City Block Distance candidate set;
The city block distance is the sum of the
lengths of the projections of the line }
segment between the points onto the delete jobi from
coordinate axes. candidate set;
}
2. RECOMMENDATION SYSTEM OF
IV. EXPERIMENTS
STUDENTS JOB
HUNTING(SJH) In this section, user-based and item-based
A. Procedure of SJH Recommendation CF algorithms are tested on our data set
There are four steps in our system as respectively. Then item-based CF,the better
Fig. 1 shows: performed one, to be applied on the Student
JobHunting recommendation system. At
last, we evaluate the performance of
improved recommender that using
usedliked job and co-apply users weights
based on item-based algorithm. The
implementation of our experiments is based
on ApacheMahout.

B. Item-based CF Deals with

Boolean Data:

Rather than using similarities between

users’ rating behavior to predict
preferences, item– based CF uses
similarities between the rating patterns of

45
International Research Journal of Engineering and Technology (IRJET) e-ISSN:
2395 -0056
Volume: 04 Issue: 05 | May -2017 www.irjet.net
p-ISSN: 2395-0072

Fig. 2 shows that all three user-based

algorithm and item basedalgorithm which
using Log likelihood similarityreached
higher precision and recall than other two
algorithms.

Under this circumstance, we continue to

evaluate these four methods via some other
variables, for example the number of
neighborhood.
C. Evaluation Of User-based CF Figure. 4. Recall of user-based CF with
1) Contrast of different similarities different similarity
methods
Since user’s preference on jobs values , in
the job apply records, and Log likelihood,
It should be noted that the Generic
City Block and Tanimotoare three methods
Recommender IR Stats Evaluator that
of similarity calculation that are suitable
Mahout offered just focuses on users who
for Boolean data, in this experiment we
have recommended items, while
recommended three items to test precision
automatically ignores users that cannot be
and recall with these three methods. We
recommended. So the result of evaluation
chose different neighborhood numbers to
seemed very good. In next section we will
reduce its influence.
test the recommendation results artificially.

TABLE I. PRECISION OF USER-

BASED CF WITH D. Evaluation of Item-based CF

DIFFERENT SIMILARITY METHODS According to the result of section IV.C,

when using Log likelihood similarity
method, item-based algorithm

similarity methods

Figure. 5. Performance of improved

recommender in
SJH system(r_num=3)

46
International Research Journal of Engineering and Technology (IRJET) e-ISSN:
2395 -0056
Volume: 04 Issue: 05 | May -2017 www.irjet.net
p-ISSN: 2395-0072

VI. PERFORMANCE OF recommended jobs, the improved

IMPROVED RECOMMENDER IN recommender changed the order of these
SJH three jobs, so that the top 2 jobs are different
from former ones. The increased scores
SYSTEM (R_NUM=2) suggest that the improved recommender
works well for the reason that jobs take
precedence (1st and 2nd recommended jobs)
are better than latter ones (3rd recommended
job).

V. CONCLUSION

On the basis of this study and various

techniques to research and after
implementation of algorithms the CF based
algorithm for its better performance and
overall factors. Of course a lot of
improvement and hybrid algorithms need to
be implemented alongside CF algorithm. To
further optimize the recommendation
As Fig.5 showed, when recommending three system, and integrate the system for better
jobs for each student, the improved performance we keep in check the sparsity of
recommender had a little promotion at user profile and use some methods of filling
Precision, recall and F1 score. When the user’s preference matrix can be utilized.
number of recommended item came to two,
as Fig.6 showed, all these three indicator
score increased significantly. And the Reach VI.REFERENCES
Rate remains as before. Because the
sparseness of our apply records dataset, [1] Pazzani M J, Billsus D. Content-
recommender can only offer 3 or based recommendation systems
4recommended results for some students. To [M]//The adaptive web. Springer
evaluate the overall recommender’s Berlin Heidelberg, 2007: 325-341.
performance, we considered the number of [2] Adomavicius G, Tuzhilin A. Toward
recommended items as three and two. If user the next generation of recommender
U has three recommended jobs---(job 1, job systems: A survey of the state-of-the-
2, job 3), when recommending jobs for U, art and possible extensions [J].
the improved recommender just re-ranking Knowledge and Data Engineering,
the three jobs recommended from the IEEE Transactions on, 2005, 17(6):
traditional recommender, so that it has no 734-749.I.S. Jacobs and C.P. Bean,
influence on the precision, recall and F1 Fine particles, thin films and
score. However, when evaluating the Top 2

47
International Research Journal of Engineering and Technology (IRJET) e-ISSN:
2395 -0056
Volume: 04 Issue: 05 | May -2017 www.irjet.net
p-ISSN: 2395-0072

[3] exchange anisotropy, in Magnetism,

vol. ))), G.T. Rado and (. Suhl, Eds.
[4] New York: Academic, 1963,
pp. 271-350.
[5] Schafer J B, Frankowski D,
Herlocker J, et al. Collaborative
filtering recommender
systems[M]//The adaptive web.
[6] Springer Berlin Heidelberg,
2007: 291-324.
[7] Sarwar B, Karypis G, Konstan J, et
al. Item-based collaborative
filtering recommendation
algorithms[C]//Proceedings of the
10th international conference on
World Wide Web. ACM, 2001:
285-295.

48
Artificial Intelligence and Human Resources Management: A
Bibliometric Analysis
P.R. Palos-Sánchez a , P. Baena-Luna b, A. Badicuc, and J.C. Infante-Moro d

aFinancial Economics and Operations Management, University of Seville, Seville, Spain; bBusiness Administration and
Marketing, University of Seville, Seville, Spain; cEconomics and Business, Open University of Catalonia, Barcelona, Spain;
d
Economics and Operations Management, University of Huelva, Huelva, Spain
ABSTRACT
Artificial Intelligence (AI) is increasingly present in organizations. ARTICLE HISTORY
In the specific case of Human Resource Management (HRM), AI Received 24 June 2022
has become increasingly relevant in recent years. This article Revised 1 November 2022
Accepted 4 November 2022
aims to perform a bibliometric analysis of the scientific literature
that addresses in a connected way the application and impact of KEYWORDS
AI in the field of HRM. The scientific databases consulted were Artificial intelligence; human resources management;
Bibliometrix; personnel recruitment; emerging
Web of Science and Scopus, yielding an initial number of 156 technologies
articles, of which 73 were selected for subsequent analysis. The
information was processed using the Bibliometrix tool, which
provided information on annual production, analysis of journals,
authors, documents, keywords, etc. The results obtained show
that AI applied to HRM is a developing field of study with
constant growth and a positive future vision, although it should
also be noted that it has a very specific character as a result of
the fact that most of the research is focused on the application of
AI in recruitment and selection actions, leaving aside other sub-
areas with a great potential for application.
Artificial Intelligence: A New Paradigm in Human Resource Management
The supposed “Fourth Industrial Revolution” or “Industry 4.0” has introduced intelligent
technologies like Artificial Intelligence (AI) (Kong et al. 2021). The increased development of
information and communication technologies (ICT) allows phenomena like AI to greatly
influence different parts of society (Bolander 2019) becoming one of the most relevant
elements of all possible changes in various aspects of life in this era (Aloqaily and Rawash
2022)
Although different departments of multiple organizations have adopted or integrated AI-based
tools, the Human Resources (HR) department still cannot implement them (Vrontis et al. 2022).
Despite there being many people in the HR department of organizations that recognize the
importance of applying AI, they also point out that they have not taken any actions regarding
this. This is a reality that shows that even though AI in the HR area is still a developing

CONTACT P. Baena-Luna [email protected] University

of Seville. Ramon y Cajal Av, 1. 41018, Seville Spain
© 2022 The Author(s). Published with license by Taylor & Francis Group, LLC.

49
This is an Open Access article distributed under the terms of the Creative Commons Attribution
License (http://creativecommons.org/ licenses/by/4.0/), which permits unrestricted use,
distribution, and reproduction in any medium, provided the original work is properly cited.
revolution and is mostly limited to large companies (Bolton 2018), it is already unstoppable.
Due to the relative novelty of this technology and its application in different areas of the
organization, many of the scientific developments in this field have mostly occurred in recent
years. For this reason, although AI has been presented as a powerful tool in HRM, academic
research on the subject is not very extensive (Pan et al. 2022).
In this context, we consider that based on a bibliometric approach, the article aims to identify
and analyze the connection of the AI phenomenon with the human resource management
(HRM) of organizations to study (1) the level of knowledge and training of their managers, (2)
the benefits and challenges in its implementation, and (3) identify the subareas with greater
development and implementation in HRM.
The connection between AI and HRM allows us to establish the following research questions
for this work. The first research question is related to previous AI reflections and challenges.
However, authors seem unclear how AI will affect or benefit employees and societies (Mitchell
and Brynjolfsson
2017). Other authors point out to the need for more data about on the speed of AI progress
(Nedelkoska and Quintini 2018). Especially its impact on every HRM-related task.

RQ1. Does the scientific community consider AI to be a commonly used tool in HRM?

The second proposed question has been studied by several published works that indicate the
benefits of AI technology in different HRM sub-tasks (Qamar et al. 2021).

RQ2. Does AI have a similar impact on all HRM sub-areas?

RQ3. Are employees in HR areas prepared to meet the challenges posed by AI in people
management?

RQ4. Does the application of AI in HRM help to improve the company’s competitiveness?

The answer to this research questions derived from the results obtained together with the
discussion and the most relevant conclusions support the theorization presented in this paper.
Regarding the originality of this work, this study, based on quantitative and qualitative
research, from the combined use of the most relevant scientific databases, Web of Science and
Scopus, allows us to focus on how IA has been integrated into organizations in HRM and its
influence on the
Table 1. Definitions of Artificial Intelligence.

(McCarthy 1956) The science and engineering of creating intelligent machines, especially
intelligent computer programs.

50
 (Minsky 1968) The science that deals with the development of machines capable of
performing functions that a human can perform and that require human
intelligence.
(Nilsson 1998) AI is a part of computer science that focuses on machine learning, making
computers act intelligently, continuously learning, and improving their
performance.
(Cappelli et al. 2019) Broad class of technologies that enable a computer to perform tasks that
normally require human cognition, including decision-making.
(Stanley and Aggarwal Development of computer systems that perform tasks that require
human intelligence.
2019) The main goal of AI is to make machines more intelligent.
(Bolander 2019) Construction of machines – computers or robots – that can perform tasks that
otherwise only humans have been able to do.
(Paesano 2021) Systems that exhibit intelligent behavior by analyzing their environment and
performing actions, with a certain degree of autonomy, to achieve specific
objectives.
approach of organizations and Human Resources. The results obtained will allow make the
following contributions. First it will serve the research community in the AI field and its
applications in the management of people and talent in organizations as a starting point for
future related research work. Also important will be the implications for the people responsible
by allowing the knowledge of the main uses and applications of new resources and tools in the
HRM of organizations will also be relevant and current trends in their application.

Concept of Artificial Intelligence

The concept of AI has multiple definitions. Different researchers have proposed their
definitions (Welsh 2019). Depending on the time and the level of technological development
reached, different studies have focused on its various aspects. A sample of the most relevant
definitions since the 20th century is shown in Table 1.
Despite the ambiguous origin of the concept of AI, two authors stand out in its development.
On the one hand, we have A.M. Turing, the father of modern computation, while on the other
hand, there is J. McCarthy, the father of AI. Turing (1937) introduced the concept of algorithms
and laid the foundation of computer science. Later, Turing (1950) proposed the Turing test,
which tests whether a machine has the capacity to be as intelligent as the person performing its
functions. However, J. McCarthy coined the term “artificial intelligence” during a conference
in Dartmouth (Paesano 2021). In the 1950s and 1960s, AI was expected to develop rapidly into
computers and robots with human-level cognitive capabilities, but that did not happen until it
recently gained prominence (Bolander 2019; Pillai and Sivathanu 2020).
Artificial Intelligence Applied to People Management
Human capital is a differentiating element of an organization as it is an intangible resource that
is difficult for competitors to imitate, thus giving a potential competitive advantage to any
organization (Kearney and Meynhardt 2016).
HRM has become a strategic trend in organizations due to economic, political, social, and
especially technological changes (Jatobá et al. 2019). Not all departments have embraced this
new role, and strategic positioning remains slow and sometimes problematic (Poba-

51
Nzaou et al. 2020). In these cases, incorporating technologies like AI requires the need to
evolve with the other facets of society (Michailidis 2018).
The role of AI in an organization is to improve efficiency and effectiveness of the HR function
by making the various management processes agile and accurate (Nankervis et al. 2021). For
HRM, IA will enable the understanding and control of a data collection process so that this
process is included in an organizational and economic efficiency strategy (Varma et al. 2022).
Among the different areas that make up the HRM in an organization where AI is starting are:
(1) talent search and recruitment, (2) training and development, (3) performance analysis, (4)
career development, (5) compensation, and (6) staff turnover (Abdeldayem and Aldulaimi
2020; Nawaz
2020; Qamar et al. 2021; Yahia, et al. 2021)
Qamar et al. (2021) showed that AI has been implemented in HRM in various organizations
via the following techniques:
Expert Systems: They are programs designed to configure expert knowledge into logical
structures that solve unstructured problems and help develop complete information systems by
providing easy access to knowledge. It is applied mainly in HR planning, compensation,
recruitment, and labor management (Malik et al. 2022).
Fuzzy Logic: This technique is used in different research fields (Salmerón and Palos-Sánchez
2019). In the case of HRM, it’s based on set membership levels, whose values vary between 0
and 1. A value of 0 indicates no membership, while a value of 1 shows full membership. With
these sets, fuzzy logic can quantify data uncertainty and foresee future scenarios to facilitate
decision- making (Kimseng et al. 2020). Its application began in 2000 and was used in
personnel selection and optimal workforce design (Qamar et al. 2021).
Artificial Neural Networks: This application is a simplified model developed to mimic the
function of a human brain. Its structure comprises a processing element, a layer, and a network
to recreate the human learning process (Huang et al. 2006). It is one of the most popular
techniques for prediction and is mainly used in selection, recruitment, and personnel
performance management (Qamar et al. 2021).
Data Mining: It is the extraction of valuable but hidden information. Through its application,
organizations can transform useful information and patterns into competitive advantages
(Huang et al. 2006). Data mining was used in HRM in 2006 and has been applied mainly for
recruitment, competency and performance evaluation, and talent management.
Genetic algorithm: These information search techniques based on replication, mutation, and
gene crossover arrive at optimal solutions to mathematical problems. It is used mainly in
workforce planning and personnel performance evaluation (Zhang et al. 2021).
Machine learning: It is the learning process by which a machine can learn by itself without
being particularly programmed to do so (Rąb-Kettler and Lehnervp 2019). Several papers agree
that the use of machine learning in decision-making is quite beneficial for HR managers and
turnover prediction (Hamilton and Davison 2022).

52
Benefits and Challenges of Artificial Intelligence in Human Resources Management
As with any technological advance, AI brings both benefits and challenges, and its application
in HRM is no different (Vrontis et al. 2022). These can be approached from three points of
view: employees, company, and society.
We highlight the following potential benefits:
Employees: The automation of repetitive and time-consuming tasks allows HR managers to
focus on those tasks that add value and require unique skills and abilities (Pillai and Sivathanu
2020). The reduction or minimization of errors owing to machine learning also helps improve
decision-making by providing more and better-processed information (Michailidis 2018).
According to a 2019 survey, 61% of companies were using AI to improve HRM in key AI-
transformed HRM areas. This task will include time-consuming and labor- intensive processes
in recruitment, such as reading many CVs, sorting through them and identifying the best
candidates and detect employees who need some training (Rykun 2019)
Company: For companies, AI means greater effectiveness and efficiency as it streamlines
management processes and reduces associated costs (Nankervis et al. 2021). It enables greater
candidate outreach as it reaches passive candidates who are not in active job search but might
become interested in the position (Black and van Esch 2021). Another important element for
companies is the improvement of communication and interaction possibilities among
employees (Michailidis 2018). Research articles looks at how AI help to improve the
successive stages of the recruitment process: identifying, selecting and retaining talented
people (Allal-Chérif et al. 2021).
Society: The creation of new professional profiles linked to AI, like robotics specialists, data
scientists, deep learning experts, generate new scenarios which can benefit the public
(Michailidis 2018).
As far as challenges are concerned, the following can be highlighted:
Employees: The application of AI may contribute to burnout, with some employees being
worried about their career uncertainty, since machines may replace them, thereby creating
anxiety and job insecurity (Kong et al. 2021). There is also dehumanization of personal
relationships, as some of the HRM processes may be performed entirely by machines, like the
use of chatbots (Fritts and Cabrera 2021). This implies the continuous need for training in
technological matters. Finally, it is necessary to point out that the “techno- stress” is a
consequence of excessive and continuous use of any type of technology (Malik et al. 2021).
Company: The need for highly qualified personnel to manage and acquire the necessary skills
to keep up with the increasing technological developments (Abdeldayem and Aldulaimi 2020)
is a reality in AI. Even though it has high implementation costs, it can reduce costs in the
processes where they are applied (Michailidis 2018). Another challenge is the existence of
biases due to the use of small and non-representative data volumes (Soleimani et al. 2022) and
the increased exposure of the company leading to increased risk of its data security breach
(Malik et al. 2021).
Society: One of the main challenges in this area is the “technology gap” Since technology in
general and AI has divided the world, it has created greater technological inequality. This is
because not all countries can implement and maintain technological infrastructure

53
(Abdeldayem and Aldulaimi 2020). Potential job losses in certain professions are also
important in the face of these challenges (Hamilton and Davison 2022).
Methodology
The methodology used was bibliometric analysis using the Bibliometrix application. This tool
was developed by Aria and Cuccurullo (2017) to carry out comprehensive analyses of the
scientific mapping of a topic. It is an open-source tool to perform a comprehensive analysis of
the scientific literature. It was programmed in R language to be flexible and facilitate
integration with other statistical and graphical packages. Bibliometrix enables the structured
analysis of large amounts of information to infer: (a) trends over time, (b) which topics are
being investigated, (c) changes in the boundaries of disciplines, etc., thus summarizing a topic
(Guleria and Kaur 2021).
The first step was to determine the databases to be used for the document search. The databases
being queried were Web of Science and Scopus, as they are currently the most relevant within
our research field (Parris and Peachey 2013). The search keywords on both bases were
“Artificial Intelligence” and “Human Resources” in the search field (Article Title, Abstract,
and Keywords) (Macke and Genari 2019) for the period 2018–2022. This period was chosen
based in previous authors. For Kshetri (2021) AI-based HRM applications can bring about
significant changes in human resource management practices. However, previous researchers
have observed a substantial gap between the promise and reality of AI in HRM (Michailidis
2018; Tambe et al. 2019). The research domain of AI in HRM is relatively nascent (Strohmeier
and Piazza 2013). Garg et al. (2022) note the narrowing of the gap between the number of
journal and conference papers from 2017 onwards: a decrease in conference papers with a
simultaneous increase in journal papers shows the increasing confidence, interest, and
acceptance for AI, especially Machine Learning (ML).
Based on all this, the choice was justified because (1) it had the highest number of publications
on the issue, (2) it had an interest in the topic, and (3) the previous literature does not correspond
to the current technological level.
Subsequently, the scientific fields selected for the query were (1) Business, (2) Management
and Accounting, (3) Arts and Humanities, (4) Social Sciences, (5) Economics and Finance, and
(6) Psychology and Research Management. These areas were chosen since they were directly
related to our current scenario. The scientific fields that could contribute the least to research,
such as physics, biology, medicine, etc., were eliminated. The analyzed works were those
written in English to cover a larger number of publications (Gutiérrez and Maz 2004) and
limited to those publications that were only articles (Podsakoff et al. 2005) excluding works
corresponding to the following types of documents: (a) book, (b) book chapter, (c) proceedings
paper, (d) review and (e) editorial material (Vlačić et al. 2021).
Once the search string was established and the corresponding filters applied, we obtained 156
articles. As shown in Figure 1, among the 89 articles initially obtained (63 in the Scopus
database and 26 in Web of Science), 9 were rejected after further analysis of their content. This
resulted in 80 valid articles for the study. Finally, 7 of these were eliminated as duplicates were
found. Thus, we finally obtained 73 papers due to the harmonization of the results.
The final articles were exported from the databases in their respective formats; Plain text and
BibTeX for Web of Science and Scopus, respectively. They were then integrated into a single

54
format to be imported later into the Biblioshiny platform and further data analysis was carried
out. Before processing the data with this software, the following steps were adopted: (1)
Download and install the latest version of R and RStudio (https://cran.r-pro ject.org/and
https://www.rstudio.com) (2) Open RStudio and in the console window type the following
command to finish the installation of Bibliometrix;

Figure 1. Identification of articles to be analyzed.

install. packages (“bibliometrix”) (3) Type the following command to be able to run the
Biblioshiny program: library (bibliometrix) biblioshiny ()
According Iden and Eikebrokk (2013) and to the established inclusion and exclusion criteria
the data extracted from each study were as follows: (1) the journal and full reference, (2) the
authors and their institutions, (3) the countries where they were situated, (4) the keywords, (5)
classification of the research methods, (6) theoretical frameworks and references theories used,
(7) main topic area, (8) research questions and (9) a summary of the study.
The critical examination of the content of each article (Bellucci et al. 2021) together with the
use of the Bibliometrix tool, in particular, by means of Multiple Correspondence Analysis
(MCA), made it possible to establish three thematic clusters: (1) AI in HRM, (2) Digital
Recruitment and (3) Electronic HR. According Paul et al. (2021) the systematic review of a
topic in depth and with rigor favors both the theory on an area and the research methodology
in that field can benefit, this is our purpose with the development of this work in the field of IA
and HRM.
Results
AI is undoubtedly one of the most important innovations. Both academics and practitioners
hope that IA can solve this problem and offer a solution to support and streamline innovation
processes. However, the literature on this topic is fragmented (Pietronudo et al. 2022). These
authors concluded that AI renews the organization of innovation and AI triggers new
challenges. That is, they suggest that AI is not a tool that uniformly optimizes innovation
management and decision-making but is better understood as a multifaceted solution.
Similar conclusions can be reached by first analyzing other systematic literature reviews
(SLRs) and a bibliometric analysis. Table 2 shows SLR works approach these reviews from
different points of view applied to a greater or lesser extent to different HRM processes.

55
The 73 articles were published in 53 different journals and represented 199 different authors
(Table 3). The average number of annual publications represents an average of < 1 article per
year, thereby indicating, at first glance, that the field of AI being applied to HRM is
underdeveloped. However, as shown in Figure 2, it is a topic of great interest in the immediate
future.
The annual distribution of the number of articles shows the general state of research and trends,
with exponential growth occurring only in the last five years. Advances and growth in the
importance of AI in both academia and HR (Jatobá et al. 2019) have sparked increased interest
in investigating the influence of one topic on the other. Although there were only two articles
in 2017 addressing the concepts in a connected way, the number increased to 10 in 2019. The
trend line shows that AI will soon persist in the future as one of the top world innovations
(Qamar et al. 2021) with an annual growth rate: 64.38%.

Analysis of Sources

Table 4 shows, in order from the highest to the lowest number of articles, the main journals
that published on these analyzed realities. The journal with the highest number was
“International Journal of Manpower” with six articles, followed by the “International Journal
of Human Resource Management” and the “Business Horizons” with five and four articles,
respectively. The journals with the highest number of publications on these topics were journals
related to business or technology, with a focus on HR, like “Advances in Developing Human
Resources and/or Human Resources Management,” is also gradually gaining importance.
Another fundamental indicator called “Bradford’s Law” was used to analyze the main journals
and their importance in the field (Bradford 1976). This law allows researchers to access those
journals that provide the most information on a topic, thus reducing their search times (Figure
3).
Another fundamental indicator called “Source Growth” was used to analyze the main journals
and their importance in the field. This figure allows researchers to know the evolution of those
journals (see Figure 4). The journals “Ethics and Information Technology” and “International
Journal of Human Resource Management” present an important growth trend.
The most cited journals were “Business Horizons,” “International Journal of Human Resource
Management” and “International Journal of Manpower,” with 138, 87, and 12 citations,
respectively (see Table 5).
The g-index is calculated from the distribution of citations of an author’s publications, which
results in a set of articles ranked in decreasing order by the number of citations they have. The
Hirsch index (h-index) uses the set of the author’s most cited articles and the number of
citations it has received in other
Table 3. Summary of bibliographic information.
Description Results
MAIN INFORMATION
ABOUT DATA
Period Time 2017:2022

56
 Sources (Journals, 53
Books, etc)
Documents 73
Annual Growth Rate % 64.38
Document Average Age 1.33
Average citations per 6.699
doc
References 2511
DOCUMENT
CONTENTS
Keywords Plus (ID) 220
Author’s Keywords 304
(DE)
AUTHORS
Authors 199
Authors of single- 13
authored docs
AUTHORS
COLLABORATION
Single-authored docs 14
Co-Authors per Doc 2.93
International co- 8.219
authorships %
DOCUMENT TYPES
article 66
article; early access 4
review 3
publications. The m-index is defined as H/n, where h is the h-index and n is the number of
years elapsed since the scientist’s first publication (Aria and Cuccurullo 2017).

Analysis of Authors

Out of 199 authors, 187 published one article, nine published two articles and three published
three articles. The authors J. Black, N. Nawaz and P. van Esch stand out (Table 6). K. Chaitanya
and V. Prikshat V. start in 2021 his first article, having published a new paper every year since
their first published article.
Regarding the impact rate of these authors, Table 7 shows that once again, the authors J. Black,
N. Nawaz and P. Van Esch have the highest index (h-index of 2), which is double the average
of the other authors, i.e., 1.
The affiliation of the authors is diverse, as shown in Table 8. Two universities stand out from
the rest, i.e., University of Turin, Kingdom University and Auckland University of Technology,
with the highest number of articles of four, three and two each, respectively.

57
The publication of articles from diverse countries reflects the subject’s global importance.
Seventeen countries published papers related to the application of AI in HRM (Table 9). The
countries with the highest number of publications were the USA, India and China, with 13, 9
and 8 articles, respectively.
Although New Zealand is not the country with the highest scientific research output, it stands
out after USA (Table 10), because its articles have

Figure 2. Annual evolution of publications.

Table 4. Sources with the largest number of related publications.

Sources Articles
International Journal of Manpower 6
International Journal of Human Resource 5
Management
Business Horizons 4
International Journal of Technology 3
Management
Journal of Management Information and 3
Decision Sciences
Advances in Developing Human 2
Resources
Computers in Human Behavior 2
Ethics and Information Technology 2
International Journal of Scientific and 2
Technology Research
Asia Pacific Journal of Human Resources 1
Benchmarking 1
BPA Applied Psychology Bulletin 1
California Management Review 1
Cyprus Review 1
Employee Responsibilities and Rights 1
Journal
European Journal of Information Systems 1

58
 Forum Scientiae Oeconomia 1
Frontiers in Psychology 1
Human Resource Management 1
Human Resource Management Review 1
the highest number of citations and therefore have the highest impact on the related scientific
literature.

Analysis of Documents

The purpose of this study was to identify the most relevant and cited articles. Table 11 shows
how Dabirian et al. (2017) is the article with the highest number of total citations. It has been
cited 89 times, with an average annual citation rate of 14.83 citations. These authors argue that
as employees use information technologies to openly share and access work-related
experiences across organizations,

Figure 3. Bradford Law.

Figure 4. Source Growth.

59
their expectations and evaluations of workplaces change. Using a data collection of 38,000
reviews of the best and worst rated employers on Glassdoor, a crowdsourced online employer
branding platform, they concluded that employers could use AI to become great workplaces
that attract highly skilled employees.
Spectroscopic analysis: According to Marx et al. (2014), Reference Publication Year
Spectroscopy (RPYS) is a quantitative method for identifying the historical origins of a
research field. It creates a temporal profile of cited references for a set of papers, thus
highlighting the period in which relatively
Table 5. Impact of sources.
Sources h_indexx g_indexx m_index TC NP PY_start
Business Horizons 3 4 0.5 138 4 2017
International Journal of 3 5 3 87 5 2022
Human Resource
Management
International Journal of 2 3 1 12 6 2021
Manpower
International Journal of 2 2 0.667 14 2 2020
Scientific and
Technology Research
Journal of Management 2 2 0.667 4 3 2020
Information and
Decision Sciences
Advances in Developing 1 1 0.333 2 2 2020
Human Resources
Asia Pacific Journal of 1 1 0.5 4 1 2021
Human Resources
Benchmarking 1 1 0.333 12 1 2020
California Management 1 1 0.25 61 1 2019
Review
Computers in Human 1 2 0.25 25 2 2019
Behavior
TC: Total citations. PY_start: Year of publication start

Table 6. Relevant authors.

Authors (>=2 articles) Articles Articles
Fractionalized
Black J. 3 1.50
Nawaz N. 3 2.33
van Esch. P. 3 1.50
Avrahami D. 2 0.50
Chaitanya K. 2 0.50
Chiappetta J. C. 2 0.45
Mcneese N. 2 0.45
Pessach D. 2 0.50

60
 Prikshat V. 2 0.45
Schelble B. 2 0.45
Singer G. 2 0.50
Wang X. 2 0.67
Table 7. Author impact factor.
Authors h_index g_index m_index TC NP PY_start
Black J. 2 3 0.5 49 3 2019
Nawaz N. 2 2 0.667 4 3 2020
van Esch. P. 2 3 0.5 49 3 2019
Сhulanova O. 1 1 0.25 5 1 2019
Abdeldayem 1 1 0.333 11 1 2020
M.
Aggarwal V. 1 1 0.25 2 1 2019
Agrawal R. 1 1 0.5 1 1 2021
Aich A. 1 1 0.333 3 1 2020
Akar C. 1 1 1 2 1 2022
Akshay P. 1 1 0.333 1 1 2020
Table 8. Authors affiliations.
Affiliation Articles
University of Turin 4
Kingdom University 3
Auckland Univ. Technol. 2
Clemson University 2
Coventry University 2
Lomonosov Moscow State 2
University
Neoma Business School 2
Shandong University 2
University of Nicosia 2
University of Reading 2
Table 9. Scientific production by
countries.
Country Freq
USA 13
India 9
China 8
UK 4
New Zealand 3
Brazil 2
Denmark 2
France 2
Italy 2
Australia 1
Table 10. Average number of citations of articles by country.

61
 Country TC Average Article
Citations
USA 67 9.57
New Zealand 49 16.33
China 38 5.43
Italy 29 9.67
India 15 3.75
Denmark 7 3.50
Cyprus 6 6.00
UK 5 1.67
France 1 1.00
Poland 1 0.50
significant findings were published along with the temporal roots of a discipline.
Figure 5 shows how AI as a technology and its use in HRM has evolved and its interest in
publishing-related work has increased. Spectroscopic analysis began in 1980, when related
publications started appearing, although as depicted in Figure 5, it was not until 2017 that there
was a notable increase in the related scientific production.
The first upturn occurred in 2000, and disruptive technology gained widespread importance
during the early 2000s. Ever since then, changes have been observed in how organizations
operate and how HRs are managed (Minbaeva 2021). Until the early 1980s, 70–90% of the
company’s value was linked to tangible assets. However, since 2000, the value linked to
intangible assets has increased to 65%, with people being the “cogs in the wheel of intangible
assets” (Black and van Esch 2021). Two AI techniques are being used in HRM: fuzzy logic
and artificial neural networks, both of which aid in the optimal workforce design and
performance management.
The second upturn occurred in 2006 when knowledge management became a field of greater
importance even though it was already being studied. Since intangible factors had already
become more important, there was a greater need for HRM to obtain competitive advantages.
Using data mining will be the
Table 11. Most cited articles.
Article (Authors/Journal) Total TC Per
Citations Year
(Dabirian et al. 2017)/Business Horizons 89 14.83
(Tambe et al. 2019)/California Management 61 15.25
Review
(Vrontis et al. 2022)/The International Journal of 51 51.00
Human R. M.
(Caputo et al. 2019)/Management Decision 29 7.25
(Black and van Esch 2020)/Business Horizons 26 8.67
(Suen et al. 2019/ Computers in Human Behavior 24 6.00
(van Esch and Black 2019)/Business Horizons 21 5.25
(Gupta et al. 2018)/Journal of Information 17 3.40
Technology Teach Classes

62
 (Malik et al. 2022)/The International Journal of 16 16.00
Human R. M.
(Giermindl et al. 2022)/European Journal of 14 14.00
Information Systems
(Pan et al. 2022)/The International Journal of 14 14
Human Resource M.
(Pillai and Sivathanu 2020)/Benchmarking 12 4
(Abdeldayem and Aldulaimi 2020)/Internatio. J. 11 3.67
of Scientific & T. R.
(Kong et al. 2021)/International Journal of 9 4.5
Contemporary H. M.
(Arslan et al. 2022)/International Journal of 9 9
Manpower
(Ogbeibu et al. 2022)/Journal of Intellectual 9 9
Capital
(Michailidis 2018)/Cyprus Review 6 1.2
(Vinichenko et al. 2019)/International Journal of 5 1.25
Recent T. and E.
(Sahota 2019)/IEEE Engineering Management 5 1.25
Review
(Boustani 2022)/Journal of Asia Business Studies 5 5

Figure 5. Annual Spectroscopic Analysis of publications.

key to correctly assessing competencies and performance. Through these evaluations, it will
promote the exchange of knowledge among employees, along with the generation of new ideas
and business opportunities.
Finally, the greatest upturn occurred in 2018, since it is from this year that the study on AI
being applied to HRM began gaining importance. The endless possibilities of AI automation
generate interest in its application in HRM (Jatobá et al. 2019).

63
Keyword analysis: Keywords are essential for a bibliographic search. Their identification and
analysis are crucial for gaining in-depth knowledge of the articles’ content and the topics being
analyzed.
The most impactful frequent keywords related to AI application in HRM are AI, HR,
Management, and Machine Learning. The importance of the AI concept stands out, but to a
lesser extent than that of HRM. AI is experiencing an increase in its application in various
fields, but as far as HRM is concerned, it has not yet occurred completely.

Knowledge Structures Analysis

Conceptual structure: It refers to what the science is about, the main themes, and trends.
Specifically, multiple correspondence analysis (MCA) helps analyze categorical data to reduce
large sets of variables into smaller sets to synthesize the information in the data (Mori et al.
2014). To do this, the data are compressed into a low-dimensional space to form a dimensional
or three-dimensional graph that uses planar distance to reflect the similarity between keywords.
Three clusters or groups of content are highlighted:

(1) Cluster 1 (AI in HRM): In this first cluster, the AI tools being applied in HRM are
addressed to highlight big data and machine learning. With big data, this might support
decision-making processes, since large amounts of varied data from various sources can
be quickly analyzed, resulting in a stream of actionable knowledge (Caputo et al. 2019).
As for machine learning, the last decade has accelerated its use and applicability owing
to the availability and variety of data (Hamilton and Davison 2022). This type of learning
provides systems with the ability to learn (Soleimani et al. 2022) and mimic human skills
(Bolander 2019). Machine learning can learn from the current context and generalize
what it has learned to a new context. There are many organizations that, despite not
comprehensively using AI in HRM, use this type of algorithm (Nankervis et al. 2021).
(2) Cluster 2 (Digital Recruitment): It is the use of ICTs to attract potential candidates, keep
them interested in the organization during the selection processes, and influence their
employment choice decisions (Johnson et al. 2021). Pillai and Sivathanu (2020) point
out how talent acquisition has become a crucial function for HR managers, with
organizations going to great lengths to attract the best talent.

For van Esch and Black (2019), talent acquisition has changed from a tactical HR activity to a
business priority. The basis of competitive advantage has shifted from tangible assets to
intangible assets, thereby increasing the strategic importance of human capital to make it the
key driver. The shortage of talent in the labor market has intensified the need for human capital.
The traditional method of searching for candidates used to be a slow and costly process.
However, today, due to technological advances and digital recruitment, it is much easier and
cheaper. Furthermore, since nowadays most of society is spending increasing time in the digital
space, if companies want to attract and recruit talent, they have to do it in that space (Black and
van Esch 2021).
3) Cluster 3 (Electronic HR): This cluster presents a much more “futuristic” vision of HR which
involves complete digitalization and the use of robots in daily functions.

64
While electronic HR management stands out in using technology to facilitate HRM processes
like, recruitment, selection, training, performance management, human resource planning,
compensation, etc. (Johnson et al. 2021). Through ICTs, it is possible to achieve better control
of performance and over the employees’ behavior for greater strategic and effective
management.
Using robots in HRM also stands out. Future forecasts are that in 20 years, robots will be in
charge of making some analytical decisions that are now being made by human managers,
while humans will continue to be in charge of tasks like creativity (Stanley and Aggarwal
2019).
Social structure: It shows how authors or countries are related in a research field; the most
commonly used is the co-authorship network (Aria and Cuccurullo 2017). The authors who
stand out for having the highest number of shared publications are Black & van Esch and
McNeese & Schelble. In general terms, there is a high degree of cooperation between authors
in the publication of articles, and very few publications are being written by a single author.
In terms of collaboration between countries, the USA is the country with the highest number
of collaborations. Whereas with New Zealand and with France, it should be noted that
collaboration with the first country is much greater than with the second. Also, there are other
collaborations between Brazil and Portugal, China and the United Kingdom, Germany, and
Norway- Tunisia.

Discussion
The research questions initially raised after the results were obtained and the studies analyzed
can be answered as follows.
Q1: AI is not yet commonly used in HRM. However, its use has acquired greater relevance in
the last five years, with 2021 being the year with the highest number of publications. Authors
like Cappelli et al. (2019) assert that the application of AI in HRM has not advanced as
expected. Among the main barriers are: the complexity of HR phenomena, associated data
challenges, equity and legal constraints, and employee reactions. Poba- Nzaou et al. (2020)
states that even though the “Fourth Industrial Revolution” again highlighted the need for people
to be at the center of organizations, it seems that HR departments remain unprepared to take
advantage of this new opportunity. Nankervis et al. (2021) point out that as technology
advances, it will be impossible for the traditional HRM approach to not advance as well; in
fact, the forecast is that over the next decade there will undergo a significant change. However,
any research article indicates that social entrepreneurship will use the opportunities of Industry
4.0 to optimize its processes until 2030, but will decline complete automation, using human
intellect and AI at the same time (Popkova and Sergi 2020)
Q2: The results obtained show that the literature has largely focused on the analysis of the
application of AI in personnel selection. Qamar et al. (2021) pointed out that, although AI is
becoming increasingly important for HR, instead of trying to take advantage of this tool to
apply it to the entire people management process, they focused only on a specific sub-area. It
is meaningless to attract the best talent if you don’t have the tools to manage it. As highlighted
by Nankervis et al. (2021), the automation of certain complex processes will require
increasingly highly trained and qualified personnel.

65
Q3. The reviewed literature highlights that most employees still do not welcome the application
of AI in HRM. Nankervis et al. (2021) show that many HR professionals lack the necessary
skills and competencies to meet the challenges of AI application in HR processes, hence their
possible contrary attitude. Fritts and Cabrera (2021) highlighted the concern of HR
professionals against the use of recruitment algorithms, as they can dehumanize the hiring
process. Vinichenko et al. (2019) highlighted how many employees lacked confidence in the
integrated use of machines in the management processes because they feared being replaced
by machines. However, this is unlikely, since even if some tasks are fully automated, the human
factor will not disappear completely (Johnson et al. 2021; Kong et al. 2021).
Q4. A company gaining a competitive advantage involves several factors like customer
satisfaction through quality service, cost optimization, innovation, productivity, etc. The
primary function of any technology, specifically AI, is to improve the efficiency and
effectiveness of the HR function to help make recruitment, retention, and management easier
and more accurate, automate repetitive tasks, and reduce labor costs (Nankervis et al. 2021).
The innovation processes is a strategical practice in business companies (Bonilla-Chaves and
Palos-Sánchez 2022)
All this will result in an innovative organization full of talent with high labor welfare, which
will provide quality service to customers, obtain customer satisfaction, and lead to higher
productivity. It has also been observed that the use of AI helps predict staff turnover to avoid
the reduction in productivity derived from it.

Theoretical Contributions

Our study contributes significantly to the literature on IA and HRM implications. It is

noteworthy how we introduce the framework of previous research on AI and HRM. Through
the results obtained by applying the methodology of bibliometric analysis and systematic
review of the literature, it has been possible to ascertain the relative and insufficient attention
by the academy to these two phenomena together.
Given the lack of similar studies applying bibliometric analysis in this field of study, it can be
the first starting point on the same. This will help future researchers as a reference point for
expanding and developing the content of this study. It can also be useful for those in HR who
want to investigate and learn more about the subject and analyze the current situation to have
a minimum number of references in case they want to enter this world.

Practical Contributions

Also important are the practical implications derived from the results of this work for the
management and administration of organizations, specifically for HRM. The results obtained
provide some very important ideas that can be of great use to HR managers and experts related
to the area to understand what the main behaviors and trends have been so far when companies
adopt HRM connected to IA.

66
As noted Pan et al. (2022), an important fact is a need for managers of organizations to
encourage the development and implementation of specific resources in the field of AI in such
a way that the adoption of AI in the company is favored.

Limitations and Future Research

This research work, like others, is also subject to a series of limitations. The main limitation
has been marked by the dispersion of information and, sometimes, limited to particular issues
that do not favor a general view of the topics in a connected way.

Research Agenda

Regarding the main lines of research derived from this work, it is important to highlight the
relevance of conducting studies that focus not only on the application of AI in the recruitment
and selection of personnel, but also on the rest of the areas in HR management. It would also
be opportune to conduct studies that analyze AI’s effect on HRM in the employees of
organizations.

Conclusions
The most relevant conclusions derived from the results obtained and their analyses are:
First, there has been an extraordinary development in technology in recent years, especially AI.
Despite its development, importance of its impact in the HRM field has not been as expected.
AI application in HRM is a very specific field of study, since most of the research has focused
on its application in the recruitment and selection of personnel, besides important functions
like training, development, or personnel rotation. There is indeed an increasing interest in talent
and the recruitment of highly qualified personnel, which is necessary for facing the changing
environment and high competition. But it should be noted that talent must not only be found,
but also maintained and developed to turn it into a competitive advantage. For this reason, it is
essential to use AI technologies in other functions and extract the maximum added value from
each process.
Second, based on the results obtained, it can be seen that there are still fears and negative
feelings in HR employees and managers about the AI application. These feelings can
complicate or slow down the use of AI in this area. Although technology has strongly disrupted
the labor market and has helped create new businesses and develop existing ones, it has also
eliminated many others, thus causing greater concern. But it should be noted that AI
technologies need people for their proper management. Despite being faster, working 24 hours
a day, optimizing time and tasks, etc., AI does not have the essential soft skills for any work
environment.
Like any new technology, AI has its strengths and weaknesses. This makes it essential for HR
departments to carry out an effective AI implementation strategy to integrate it safely within
organizations, thus eliminating the potential damage. It is obvious that in the long term, the use
of disruptive technologies will no longer be optional but rather necessary to remain competitive

67
among other organizations; otherwise, they will lose their market positions or worse, will
disappear.

Disclosure Statement
No potential conflict of interest was reported by the author(s).

ORCID
P.R. Palos-Sánchez http://orcid.org/0000-0001-9966-0698
P. Baena-Luna http://orcid.org/0000-0002-8509-0222
J.C. Infante-Moro http://orcid.org/0000-0003-0239-5053

References
Abdeldayem, M. M., and S. H. Aldulaimi. 2020. Trends and opportunities of artificial
intelligence in human resource management: Aspirations for public sector in Bahrain.
International Journal of Scientific and Technology Research 9 (1):3867–71.
Allal-Chérif, O., A. Alba, and R. Castaño. 2021. Intelligent recruitment: How to identify,
select, and retain talents from around the world using Artificial Intelligence. Technological
Forecasting and Social Change 169:120822.
Aloqaily, A., and H. N. Rawash. 2022. The application reality of Artificial Intelligence and its
impact on the administrative human resources processes. Journal of Positive School
Psychology 6 (5):3520–29.
Aria, M., and C. Cuccurullo. 2017. Bibliometrix: An R-tool for comprehensive science
mapping analysis. Journal of Informetrics 11 (4):959–75. doi:10.1016/j.joi.2017.08.007.
Arslan, A., C. Cooper, Z. Khan, I. Golgeci, and I. Ali. 2022. Artificial Intelligence and human
workers interaction at team level: A conceptual assessment of the challenges and potential
HRM strategies. International Journal of Manpower 43 (1):75–88.
Basu, S., M. Bishakha, M. Kajari, M. Surender, and P. Chandan. 2022 March. Artificial
Intelligence–HRM interactions and outcomes: A systematic review and causal
configurational explanation. Human Resource Management Review 100893.
doi:10.1016/j.hrmr.2022. 100893.
Bellucci, M., G. Marzi, B. Orlando, and F. Ciampi. 2021. Journal of intellectual capital: A
review of emerging themes and future trends. Journal of Intellectual Capital 22 (4):744–67.
doi:10. 1108/JIC-10-2019-0239.
Bhatt, P. K., and A. Muduli. 2022. Artificial intelligence in learning and development: A
systematic literature review. European Journal of Training and Development 248781691.
doi:10.1108/ejtd-09-2021-0143.
Bilan, S., P. Šuler, O. Skrynnyk, E. Krajňáková, and T. Vasilyeva. 2022. Systematic
bibliometric review of Artificial Intelligence technology in organizational management,
development, change and culture. Artificial Intelligence technology in organizational
management, development, change and culture. Business: Theory and Practice 23 (1):1–13.
doi:10.3846/btp. 2022.13204.
Black, J. S., and P. van Esch. 2020. AI-enabled recruiting: What is it and how should a manager
use it? Business Horizons 63 (2):215–22.

68
Black, J. S., and P. van Esch. 2021. AI-enabled recruiting in the war for talent. Business
Horizons 64 (4):513–24. doi:10.1016/j.bushor.2021.02.015.
Bolander, T. 2019. What do we loose when machines take the decisions? Journal of
Management and Governance 23 (4):849–67. doi:10.1007/s10997-019-09493-x.
Bolton, R. 2018. The future of HR 2019: In the know or in the No. KPMG International. 1–24
.
Bonilla-Chaves, E. F., and P. R. Palos-Sánchez. 2022 A. Mesquita, A. Abreu, & J. V. Carvalho.
Perspectives and Trends in Education and Technology. Smart Innovation, Systems and
Technologies 256: 607–20. doi:10.1007/978-981-16-5063-5_50.
Boustani, N. M. 2022. Artificial intelligence impact on banks clients and employees in an Asian
developing country. Journal of Asia Business Studies 16 (2):267–78. doi:10.1108/JABS-09-
2020-0376.
Bradford, M. M. 1976. A rapid and sensitive method for the quantitation of microgram
quantities of protein utilizing the principle of protein-dye binding. Analytical Biochemistry
72:248–54. doi:10.1006/abio.1976.9999.
Briner, R. B., and D. Denyer. 2012 Systematic Review and Evidence Synthesis as a Practice
and Scholarship Tool Rousseau, D.M. Handbook of Evidence-Based Management:
Companies, Classrooms and Research (Oxford: Oxford University Press) 112–29.
doi:10.1093/oxfordhb/ 9780199763986.013.0007.
Cappelli, P., P. Tambe, and V. Yakubovich. 2019. Artificial Intelligence in human resources
management: Challenges and a path forward. California Management Review 61 (4):15–42.
doi:10.1177/0008125619867910.
Caputo, F., V. Cillo, E. Candelo, and Y. Liu. 2019. Innovating through digital revolution.
Management Decision 57 (8):2032–51. doi:10.1108/MD-07-2018-0833.
Coron, C. 2022. Quantifying Human Resource Management: A literature review. Personnel Review 51 (4):1386–409.
doi:10.1108/PR-05-2020-0322.
Crossan, M. M., and M. Apaydin. 2010. A multi-dimensional framework of organizational innovation: A systematic
review of the literature. Journal of Management Studies 47 (6):1154–91. doi:10.1111/j.1467-6486.2009.00880.x.
Dabirian, A., J. Kietzmann, and H. Diba. 2017. A great place to work!? Understanding crowdsourced employer branding.
Business Horizons 60 (2):197–205.
Denyer, D., and D. Tranfield. 2009. Producing a Systematic Review D. A. Buchanan & A. Bryman . In The Sage Handbook
of Organizational Research Methods, 671–89. Thousand Oaks: Sage Publications Ltd.
Di Vaio, A., R. Palladino, R. Hassan, and O. Escobar. 2020. Artificial Intelligence and business models in the sustainable
development goals perspective: A systematic literature review. Journal of Business Research 121 (December):283–
314. doi:10.1016/j.jbusres.2020.08.019.
Fritts, M., and F. Cabrera. 2021. AI recruitment algorithms and the dehumanization problem. Ethics and Information
Technology 23 (4):791–801. doi:10.1007/s10676-021-09615-w.
Garg, S., S. Sinha, A. Kumar, and M. Mani. 2022. A review of machine learning applications in Human Resource
Management. International Journal of Productivity and Performance Management 71 (5):1590–610.
doi:10.1108/IJPPM-08-2020-0427.
Giermindl, L. M., F. Strich, O. Christ, U. Leicht-Deobald, and A. Redzepi. 2022. The dark sides of people analytics:
Reviewing the perils for organisations and employees. European Journal of Information Systems 31 (3):410–35.
doi:10.1080/0960085X.2021.1927213.
Guleria, D., and G. Kaur. 2021. Bibliometric analysis of ecopreneurship using VOSviewer and RStudio Bibliometrix,
1989–2019. Library Hi Tech 39 (4):1001–24. doi:10.1108/LHT-09- 2020-0218.
Gupta, P., S. Fernandes, and M. Jain. 2018. Automation in recruitment: A new frontier. Journal of Information
Technology Teaching Cases 8 (2):118–25. doi:10.1057/s41266-018-0042-x.
Gutiérrez, Á., and A. Maz. 2004. Cimentando un proyecto de investigación: undefined revisión de literatura. Revista
EMA 9 (1):20–37.

69
 Journal of Management
Vol. 45 No. 6, July 2019 2498 –2537
DOI: 10.1177/0149206318818718https
© The Author(s) 2019
Article reuse guidelines: sagepub.com/journals-permissions

A Systematic Review of Human Resource

Management Systems and Their Measurement
Corine Boon Deanne N. Den
Hartog
University of Amsterdam

David P. Lepak
University of Massachusetts Amherst

In the strategic human resource (HR) management literature, over the past three decades, a shared
consensus has developed that the focus should be on HR systems rather than individual HR
practices because the effects of HR practices are likely to depend on the other practices within the
system. Despite this agreement, the extent to which the fundamental assumption in the field of
interactions and synergy in the system holds true is unclear. We present a systematic review of
495 empirical studies on 516 HR systems in which we analyze the development of HR systems
research over time and identify important trends, explicitly linking conceptualization and
measurement of the HR system. Our findings suggest that the increasingly broad conceptualization
and measurement of HR systems and the lack of clarity on the HR systems construct at different
levels have hampered research progress. Much of the research to date does not align with the
fundamental assumption of synergies between HR practices in a system, the measures have
problems and increasingly confound HR systems with related concepts and outcomes, and
insufficient attention is paid to the HR system construct at different levels. Overall, we thus still
know little about the “systems” element and how synergies and interactions in an HR system
operate. We offer actionable suggestions on how to advance HR systems research towards con -

Acknowledgments: We would like to thank John Delery, Robert Verburg, and the
participants of research seminars at the University of Southern Australia and the
Amsterdam Business School for their very helpful comments and suggestions and
Taylor Geiger and Candy Sin Man Lai for their assistance with collecting and coding
the papers. We would also like to thank David Allen and the two anonymous reviewers
for the useful and constructive feedback during the review process. Dave Lepak passed
away on December 7, 2017. He had an important role in developing the ideas in this
paper. Dave’s profound impact on the strategic HRM field will be long lasting, and his
work will continue to inspire us.
Supplemental material for this article is available with the manuscript on the JOM
website.
Corresponding author: Corine Boon, Amsterdam Business School, University of
Amsterdam, Plantage Muidergracht 12, 1018 TV Amsterdam, The Netherlands.

70
E-mail: [email protected] clarity and construct refinement, focusing both on
how to conceptualize, measure, and combine practices in systems and on studying
such systems at different levels of analysis.

Keywords: strategic human resource management; human resource

management systems; HR systems; HR bundles; synergies; internal fit;
horizontal fit; review

Strategic human resource management (SHRM) research increasingly focuses on the

performance effects of human resource (HR) systems rather than individual HR
practices (Combs, Liu, Hall, & Ketchen, 2006). Researchers tend to agree that the focus
should be on systems because employees are simultaneously exposed to an interrelated
set of HR practices rather than single practices one at a time, and the effects of HR
practices are likely to depend on the other practices within the system (Delery, 1998;
Jiang, Lepak, Han, Hong, Kim, & Winkler, 2012; Lepak, Liao, Chung, & Harden,
2006). Research indeed consistently shows a positive association between (broad) HR
systems and performance (e.g., Boselie, Dietz, & Boon, 2005; Jiang, Lepak, Ju, & Baer,
2012), and the idea of complementarities or synergies between practices in an HR
system is widely accepted as the conceptual logic behind the effectiveness of HR
systems (e.g., Chadwick, 2010; Delery, 1998; Gerhart, 2007; Jiang, Lepak, Han et al.,
2012). Despite this agreement, the extent to which this fundamental assumption in the
field of interactions and synergy in the system holds true is unclear. In other words, our
understanding of the “systems” element of HR systems seems more nascent than one
might expect, given the sizable body of literature on HR systems.
In the past, several authors have noted fundamental problems in the research relating
to how the “system” element of HR systems has been conceptualized. For example,
over a decade ago, Lepak and colleagues (2006), in a review of HR systems research,
highlighted that a wide variety of HR systems exist with labels such as high
performance, commitment, and involvement HR systems but that how these systems
are distinct in terms of the practices they include or exclude, how the selected practices
help achieve the system’s goal, and why these systems would have distinct effects on
outcomes was not sufficiently clear. Our first aim is to review the available empirical
studies on HR systems and compare studies over time to assess the extent to which the
field has progressed in dealing with these issues. In addition, despite the agreement on
the interactive nature of HR practices, no consensus has developed on how to combine
HR practices into (synergistic) systems (e.g., Chadwick, 2010), and it remains unclear
whether or how the field has progressed in terms of understanding how interactions
within HR systems that are supposed to be complementary or synergistic work. Thus,
going beyond previous reviews, our second aim is to assess the different ways to
combine practices in HR systems studies to date to address whether and if so, how the
field has progressed in assessing the synergistic effects of HR systems.

71
Construct development concerns the simultaneous process of validation of measures
and theory, and because theory and measurement are inherently linked, both need to be
considered in order to advance theory (Smith, Fischer, & Fister, 2003; Strauss & Smith,
2009). The HRfield has paid relatively little attention to measurement of HR systems,
and previous reviews have not yet focused in detail on these measures. While of course
using different measures of the same underlying construct is of value to advance theory,
if the same HR system is measured in vastly different ways without clarity as to why
this is the case, the question does become whether measures indeed still capture the
same underlying construct and, thus, whether results of such studies are sufficiently
comparable. Without good measurement and sound study design, empirical findings
may reveal more about the measure than the construct, leading to inaccurate or
misleading results (McGrath, 2005; Rossiter, 2008). Thus, our third aim is to review
the development of study design and measurement of HR systems over the past three
decades.
In sum, we present a systematic review of existing empirical studies on HR systems
and analyze the development of the field over time. We take a comprehensive approach
and focus on all choices researchers make when designing a study on HR systems,
explicitly linking conceptualization and measurement of the HR system. We analyze
developments in how HR systems have been conceptualized and measured, how
practices are combined into systems, and how HR systems studies are designed. On the
basis of this, we highlight conceptual and empirical problems in the current field and
offer practical guidance on addressing some of the limitations undermining the current
empirical literature, and we discuss theoretical and methodological advances needed to
progress towards a better understanding of HR systems.
Our review extends previous work in several important ways. First, analyzing the
development of HR systems research over time enables us to identify areas in which
progress has been made and where such progress is lacking. In doing so, we identify
the most pressing research needs and develop a future research agenda aimed at better
understanding interrelationships between HR practices in a system. Second, we add
to previous reviews through our focus on both the conceptualization and the
measurement of HR systems. Beyond prior reviews, which were primarily conceptual
with some addressing some aspects of study design, we also review HR system
measures at the item level. As noted, jointly considering both theory and measurement
is needed, and in doing so, we identify future research directions that can help
establish correspondence between conceptualization and measurement and provide a
stronger basis for further theory development on HR systems. Third, we focus
specifically on the system element of HR systems by assessing every aspect of HR
systems research. Most reviews focus either broadly on the field of SHRM and
identify important themes such as human resource management (HRM)
implementation or mediating mechanisms in the HRM–performance relationship
(e.g., Jackson, Schuler, & Jiang, 2014; Jiang & Messersmith, 2018; LengnickHall,
Lengnick-Hall, Andrade, & Drake, 2009) or on specific issues (e.g., levels of analysis,
Arthur & Boyles, 2007; Peccei & Van De Voorde, 2019; high performance work
practices, Posthuma, Campion, Masimova, & Campion, 2013). Our review is broader
in its coverage than those focused on specific issues that include the subset of articles

72
 related to that issue and is more exhaustive than those providing a broad thematic
overview that focus on a selection of impactful articles (e.g., Lengnick-Hall et al.,
2009; Wright & Ulrich, 2017).
Below, we first provide a brief overview of HR systems theory and then present our
review showing how HR systems research has developed over the past three decades.
Our findings suggest two main and interrelated issues that have hampered research
progress: the increasingly broad conceptualization and measurement of HR systems
and the lack of clarity on the HR systems construct at different levels. In addition, we
see confounding of HR systems with related constructs and outcomes. Together, these
problems imply that it is not always sufficiently clear what is responsible for found
performance effects of HR systems, which suggests that some of the current evidence
may be misleading, and that we lack knowledge about the “system” element of HR
systems. We highlight areas of conceptual and empirical confusion in the composition
and measurement of HR systems that have hindered theory building, and we offer
actionable suggestions on how to advance HR systems research.

Literature Review
Conceptualizing HR Systems
SHRM can be defined as “the pattern of planned HR deployments and activities
intended to enable an organization to achieve its goals” (Wright & McMahan, 1992:
298). Increasingly, the field has emphasized the importance of focusing on whether and
how “systems” or “bundles” of HR practices jointly help organizations achieve
strategic goals, rather than on single HR practices individually. An HR system can be
defined as a combination of HR practices “that are espoused to be internally consistent
and reinforcing to achieve some overarching results” (Lepak et al., 2006: 221).
Conceptually, these systems of HR practices—as a whole— are proposed to affect
performance-related outcomes (Delery, 1998; Wright & Boswell, 2002). Existing
evidence provides some first meta-analytic support, as HR systems tend to be more
strongly related to performance than individual HR practices (Combs et al., 2006).
However, how this joint effect occurs seems less clear. Conceptually, all practices in a
system are proposed to promote an overarching goal (e.g., Jiang, Lepak, Han, et al.,
2012); however, it is not always clear what the overarching goal is, how HR systems
are conceptualized, or how practices contribute to this goal.
Multiple conceptualizations of HR systems exist, including high performance (e.g.,
Huselid, 1995), commitment (e.g., Arthur, 1994), and involvement (e.g., Guthrie,
2001). Some scholars use general labels such as HR system or HR bundle without
indicating a dominant strategic focus, while others study targeted HR systems focused,
for example, on customer service or teamwork (Jackson et al., 2014). Different levels
can be distinguished within HR systems: HR policies represent an organization’s stated
intentions about HR practices that should be implemented, whereas HR practices
reflect the actual HR activities (Becker & Gerhart, 1996; Wright & Boswell, 2002).
Techniques are methods used within practices, such as assessment centers in selection.
One can also structure HR systems by focusing on broader types or subbundles of
practices, such as those based on the abilitymotivation-opportunity (AMO) model:
ability-enhancing practices (e.g., selection, training), motivation-enhancing practices

73
(e.g., performance management, rewards), and opportunityenhancing practices (e.g.,
participation, job design; e.g., Jiang, Lepak, Ju, & Baer, 2012). The logic for this level
of abstraction is that countless specific HR practices exist that at a broader policy level,
form conceptually similar groupings of practices.
Already over a decade ago authors lamented that a precise and consistent definition of
HR systems was lacking and that the variability across HR systems in terms of the
included practices was considerable (e.g., Lepak et al., 2006). Here we review whether
this has changed over time. We examine how systems are labeled and which practices
and subbundles they contain to determine how HR systems that are labeled differently
can be distinguished from each other and to what extent HR systems that are labeled
similarly indeed are similar in terms of the practices they include. Ambiguity regarding
the conceptual boundaries of a construct hinders knowledge accumulation, as it may
be unclear what we are speaking about when we examine or compare (specific) HR
systems (cf. Podsakoff, MacKenzie, & Podsakoff, 2016).

The System Element of HR Systems

The core assumption underlying HR systems research is that the effectiveness of an
HR practice depends on the other practices in the system (Delery, 1998). When
practices fit into a coherent system (internal/horizontal fit), they reinforce one another
and create synergies. When practices do not fit, they may detract from each other’s
effects. Thus, HR practices should be examined jointly rather than separately. Practices
in a system can relate to one another in different ways. For example, an additive
relationship assumes HR practices have independent effects and add up without
influencing each other. In contrast, in an interactive relationship, the effectiveness of a
practice depends on the presence or level of other practices. Practices may for instance
be substitutes or show positive or negative synergies (e.g., Delery, 1998).
Assuming an additive relationship between practices typically implies calculating an
HR system score by summing or averaging scores on individual practices into a scale
score or index (Delery, 1998). This approach assumes that HRM is best viewed as a
consistent system that has most impact if all practices send consistent signals about the
organization’s underlying intentions (Bowen & Ostroff, 2004). A suggested advantage
of an additive index is that it allows for different ways (i.e., different combinations of
practices) to achieve a high system score (e.g., Becker & Huselid, 1998). Yet many
disagree with the use of additive indices, as these cannot capture the assumed synergies
between practices, and advocate using methods that can capture these, such as cluster
analysis or interactions (Becker & Gerhart, 1996; Chadwick, 2010). The few studies
that compare different analytical techniques to test for synergies show that the different
techniques yield different results and represent different underlying ideas about fit
(Chadwick, 2010; Delery & Gupta, 2016). Overall, conceptual approaches to
combining differ considerably, and disagreement exists on how to combine HR
practices in a system. Knowing how the elements of an HR system interact is important
in order to study whether “systems” indeed affect intended outcomes. How much
empirical attention different ways of combining practices have received over time is
not clear; thus, we review this and analyze trends in the field over time.

74
Study Design and Measurement
Theory and measurement are inherently linked, and the absence of rigorous study
designs and valid measurement can hamper theoretical progress in the field. We thus
also review this. We assess who is used as the source to provide information on the HR
system. Early research relied mostly on a single (HR) manager to rate the system, which
has problems, such as the potentially low reliability of such single-informant designs
(e.g., Gerhart, Wright, McMahan, & Snell, 2000). However, even if multiple
respondents are used, these sources may not be the most knowledgeable about specific
practices or levels. For example, several studies focus on employee perceptions of HR
systems (e.g., Den Hartog, Boon, Verburg, & Croon, 2013; Liao, Toya, Lepak, &
Hong, 2009), which may not be suited for all research purposes, as employees might
not be able to fully evaluate HR systems, especially practices that do not pertain to
them personally or intended policies. The HR system may have different meanings at
different levels, with different problems associated with each of the levels. Thus, we
examine developments over time in the source used to rate the HR system and the
levels at which the HR system is theorized and analyzed.
In addition, we review answer scales, as disagreement exists about appropriate rating
or answer scales for capturing HR practices (Wright & Gardner, 2003). Answer scales
can be more objective, such as the percentage of employees a practice covers, or more
subjective, such as Likert-type scales indicating attitudes towards certain practices, and
these can reflect different constructs. We assess the examined outcome, which is
relevant as, for example, when studies measure how employees feel about the HR
system and relate this to attitudinal outcomes, overlap may occur between the HR
system and outcome. Also, because HR system theory implicitly assumes that time is
important, as HR systems are supposed to influence performance, the field needs study
designs that allow testing for relationships over time and cannot rely on cross-sectional
designs. Thus, we review whether longitudinal studies are done and what they focus
on.

Method
Literature Search
We conducted a search of the peer-reviewed academic literature on HR systems
published before September 2017. We searched the Scopus and OVID PsycINFO
databases, and cross-checked with the EBSCO Business Source Premier database. We
searched for peer-reviewed articles containing the following keywords in the title or
abstract: “human resource management system” (or human resource/HR/HRM
system), “HR(M) bundle,” “HR(M) configuration,” “set of HR(M) practices,” “human
resource (management) practices,” “high performance/ involvement/ commitment
work system” (or high performance/ involvement/ commitment HR/HRM/work
practices). In addition, we sent a message to the HR division listserver asking for in
press articles. Our deletion of duplicates yielded 5,303 articles. To get a representative
picture of the field, which is sufficiently comprehensive and manageable and of
sufficient quality, we focused on journals with an impact factor over 1. Thus, we
removed all articles published in journals without an impact factor (964 articles) or
with an impact factor below one (451 articles), resulting in 3,888 articles. To be

75
included, an empirical study had to meet the following criteria. First, it had to focus on
multiple HR practices. Studies on a single practice were excluded. Next, it had to use
a quantitative methodology and measure the HR system with a measurement scale.
Third, it had to combine the HR practices in some way in a system in the analyses. We
did not consider studies in which HR practices were included individually in the
analyses. In total, 495 articles met the criteria and were included in our review; these
articles are listed in the online supplemental material.

Coding Procedure
Conceptualization

Some papers report multiple studies or use multiple HR systems; thus, the 495 articles
included 516 HR systems. We coded these 516 systems using the following criteria.

HR system label. We coded the label that is used for the HR system, usually retrieved
from the hypotheses, model, and tables. Categories were unspecified (for general
labels, e.g., HR system, HR practices, HR configuration), high performance, high
commitment, high involvement, (strategically) targeted (for labels that clearly specify
the target of the HR system), and other.

HR practices or practice domains measured. On the basis of Lepak et al. (2006) and
Combs et al. (2006), we coded the following HR practices: job analysis/job design,
recruitment, selection, training and development, incentive compensation, other
compensation, (self-managed) teams, participation/autonomy, (results-oriented)
performance appraisal/ management, job security, employee voice/grievance,
promotion from within/career development/internal labor market, information
sharing/communication, HR planning, flexible work/family-friendly practices, and
other practices. We also coded how many practices were included.

Subbundles. We coded whether the study distinguishes between subsystems or

subbundles. Categories were ability bundle, motivation bundle, and opportunity bundle
(i.e., AMO model), as well as other and none. We coded only subbundles included in
the analyses as separate bundles. When subbundles were mentioned only in theory or
in discussing the overall HR system, but not included as variables in the analyses,
subbundles were not coded.

The Type of Relationship Between the Practices and Bundles

We coded how individual HR practices were combined in systems. All studies that
combined practices by averaging or summing scores of the individual practices or used
subscale aggregation were coded as additive index, and a second category included
studies that analyzed the HR system as a latent factor. All other approaches were first
listed under the category other, and subsequently this group was further coded on how
they combined practices (see the appendix). We included a category for unclear when
no information was provided. We also coded whether and how subbundles were
combined in analyses (included as separate bundles or other approaches).

76
Study Design

We coded all 495 articles in terms of their study design using the following criteria.
Levels. We coded the level of theory and level of analysis of the HR system. The level
of theory was coded as organization when theory assumed differences between
organizations or when employees were considered as one homogeneous group, as
group/unit when assuming differences between units but units being homogeneous, and
as individual when differences between individuals were assumed. Categories for level
of analysis of the HR system were organization, group/unit, and individual. We also
coded whether the study tested a multilevel model.

Data source. We coded who filled out the HR system measure: HR professionals,
higher/ middle-level managers (e.g., CEOs, unit/department managers), line (or team)
managers, employees, others, or unclear. In addition, we coded the use of one or
multiple sources.

Answer scale. Categories were presence (yes/no), coverage (the percentage of

employees covered by a practice), Likert-type scale, other (for other answer scales),
and unclear. We also coded whether one or multiple types of answer scales were used
in one measure.

Outcomes. We coded which types of outcome(s) were examined in each study:

attitudes, behaviors, performance (including different types of
individual/organizational performance, e.g., productivity or task performance), other,
or none (studies with the HR system as the outcome).

One or multiple time points. We coded whether studies were cross-sectional, used
separate measurements in time, or were longitudinal in nature.

Measures

We coded whether the measure for the HR system was existing, adapted from existing
measures, or newly developed. For the adapted ones, we listed references to the original
measures up to three, and when four or more were used, we coded them as multiple.
Of the 516 systems, 219 had (mostly) new measures, 193 adapted ones, and 100 an
existing measure. For 4 of the systems, it was unclear. Part of our review focuses on
the item level. For this, we needed full measures. For 209 studies, the measure was
available in full in the article; of these, 29 were existing, 77 were (mostly) new, and
103 were adapted from existing ones. Of these, 34 were adapted from four or more
measures. We coded the 77 newly developed ones and the 34 based on four or more
existing ones (111 in total) for the following.

Policies, practices, or techniques. Items were coded as policies if they referred to

organizational goals or objectives for managing HRs. We coded items referring to
general practices, such as selection, as practices and as techniques if they referred to
specific practice techniques used within a practice, such as selection interviews or
assessment centers.

77
General vs. criterion focused. We coded whether items were general (e.g., referring to
rigorous selection) or focused on a specific criterion (e.g., selection based on
creativity).
Who offers HR practices. Different agents can offer HR practices, and we coded
whether items referred to HR practices emanating from the organization, unit, or
manager. We used unspecified when it was unclear who offered HR.

Item referent. We included the following categories when coding item referents: group
(multiple individuals, such as employees, as the referent), job (a specific job or job
category as the referent), individual (one individual as the referent), or
unspecified/unclear.

Item focus. We coded whether items were descriptive or evaluative. When items refer
to a practice in an objective way (e.g., how many hours of training), we coded them as
descriptive, and when items contain a value judgement or refer to a feeling, we coded
them as evaluative (e.g., communication is effective). We also used the category
descriptive with Likert scale for descriptive items with a Likert scale, which includes
more evaluation than percentages or coverage. We used the category descriptive and
evaluative for mostly descriptive items that contain an evaluative element, using words
such as “considerable” or “serious” (e.g., considerable importance is placed on
staffing).

Results
Table 1 summarizes the coded data for the 516 systems on how HR systems are
conceptualized and combined, and Table 2 summarizes the coded data on study design
and measurement. To assess developments over time, we report results for five time
periods (1991–2000, 2001–2005, 2006–2010, 2011–2015, 2016–2017) and the total
period (Total).1 When reporting changes, we report percentages for the first (1991–
2000) and last (2016–2017) period.

Conceptualization of HR Systems
How Are HR Systems Labeled?
Table 1 shows that many different HR system labels are used. Unspecified labels such
as HRM, HR practices, HR system, HR bundle, or HR configuration are widely used
(34% overall), but their use has decreased over time (from 59% to 23%). With these
generic labels, it is unclear what the goal of a system is. Labels such as high
performance (35%), commitment (8%), or involvement (8%) HR systems are widely
used with little change over time. Table 1 shows that targeted HR systems with more
specific labels such as relationship-oriented HR system, knowledge-oriented HR
system, and initiative-enhancing HRM system are less common (12% overall) but have
increased over time (from 9% to 19%). The remaining studies (3%) mostly do not focus
on (the extent to) which HR practices are offered but on preferences for, motivation
for, satisfaction with, or effectiveness of HRM.

78
Problematically, different terms are often used for highly similar HR systems, which
has not improved over time. For example, while the labels of high performance and
high commitment HR systems suggest they are differentially strategically targeted HR
systems (focused on increasing performance vs. commitment), they are used
interchangeably in many studies, implying these labels have become more general than
originally intended. The practices included in and the items used to measure these
systems overlap strongly. For example, most practices are found in both

Table 1
Conceptualization of Human Resource (HR) Systems
1991– 2001– 2006– 2011– 2016–
2000 (34 2005 (61 2010 2015 2017 Total
Systems) Systems)(114 (202 (105 (516
Systems) Systems) Systems) Systems)

Discussion and Implications

We aimed to review three decades of HR systems research focusing on the “systems” element
of HR systems to identify where the field has progressed and where it has not and to provide
recommendations for moving this research forward. As noted, HR systems research overall
suggests a positive relationship between HR systems and performance. However, the findings
of this review show that the conclusion that research to date shows that HR systems are
effective may be misleading. In most studies, conceptualization and measurement do not
match the core theoretical assumption of complementarities or synergies between HR
practices in a system. Thus, while the empirical evidence so far may suggest that we can draw
the broad conclusion that “investments in some broad set of HR practices yields returns,”
which practices this entails and whether and how practices jointly affect outcomes remains
unclear. In addition, the measures used have problems and increasingly confound HR systems
with related concepts and outcomes; thus, it is not always clear whether it is indeed the HR
system causing effects. Finally, insufficient attention is paid to how differences between levels
affect the meaning of the HR system construct. Overall, this makes it unclear exactly what is
responsible for the found performance effects in HR systems research and shows we still know
little about the theorized “systems” element or how synergies and interactions in an HR
system operate.
Our review shows that despite earlier calls to study more specific and targeted systems (e.g.,
Lepak et al., 2006), approaches to measuring and combining HR practices in a system have
moved even further towards a focus on broad undifferentiated HR systems. Our findings also
show that over time, agreement in the field on how to measure HR systems has declined and
confounding has increased, and it remains unclear which (sets of) practices drive the system’s
effect at different levels. Also, despite calls to address nonadditive effects (e.g., Chadwick,
2010), the use of additive approaches to combine HR practices in a system has increased rather
than decreased recently. Research thus still provides only limited insight into the core
theoretical assumption of complementarities or synergies between HR practices. In addition,
theory on HR systems implicitly assumes that the HR system is influenced and shaped by
time. Some first studies suggest that practices indeed vary in the timing of their effects and
that effects of practices are likely to be nonlinear (e.g., Birdi et al., 2008; Piening et al., 2013),
suggesting that cross-sectional studies may (at times) yield inaccurate results. While some

79
progress has been made in showing causal effects of HR systems using additive indices,
longitudinal studies have hardly examined the “system” element of HR systems over time. As
very little explicit attention is paid to interrelationships between practices in a system over
time, our understanding of how interrelationships between practices in HR systems develop
and change is very limited.
The importance of (differences and differentiating between) levels in HR systems was noted
earlier (Arthur & Boyles, 2007), and HR systems are increasingly studied at different levels,
adding complexity to the conceptualization and measurement of HR systems. While this
implies progress in terms of moving beyond considering only the organizational level,
theorizing around HR systems at multiple levels has yet to follow suit, as even in studies
measuring at the individual level, by far most theory (95%) is still focused exclusively on the
organizational level. Misalignment between the level of the method and analyses and the level
of theory can yield artefactual results, with found relationships being inaccurate because they
do not capture meaningful variation at the right level (Klein, Dansereau, & Hall, 1994). Thus,
more specificity in theory on the HR system at different levels is essential to move the field
forward.
Over 80% of the studies use HR system measures that are new or are adapted from other
scales and that have not received extensive scale validation, so empirical evidence that
measures actually tap the intended constructs is limited (McGrath, 2005; Smith, 2005). The
item types used are increasingly mixed, resulting in ambiguous scales with heterogeneous
items that may not represent the same underlying construct (cf. Strauss & Smith, 2009). Also,
there is a general trend over time towards the use of more perceptual and evaluative
measurement: the use of individual employee respondents to rate the HR system and of
individual item referents is increasing (focusing on the respondents’ individual experience
rather than common experiences of the group), and more Likert scales and evaluative items
are being used.
Overall, the broad and heterogeneous conceptualization and measurement of HR systems and
lack of clarity in levels introduces theoretical and empirical imprecision because variation on
the construct may represent variation in any or all of its levels or dimensions (Edwards, 2001;
Smith et al., 2003; Strauss & Smith, 2009). This imprecision, which our review suggests is
generally increasing rather than decreasing, hinders further theory development on HR
systems. Theoretical progress in any field is typically characterized by construct refinement.
Over time, distinctions between dimensions often become increasingly clear and constructs
become more differentiated, and as a result, broader constructs become less useful (Edwards,
2001) and more rigorous empirical tests are necessary for scientific advancement (Schmidt &
Pohler, 2018). In HR systems research, however, rather than a trend towards more specific
theory development and related increasing precision in measurement, for example, by
differentiation between different possible targeted systems, we see a trend towards even
broader and less clear HR system constructs and operationalizations. From our analysis, we
signal two main and interrelated areas that need specific attention in future work on HR
systems to move the field forward in terms of construct refinement and building more
knowledge on how HR practices combined in “systems” affect outcomes: measuring and
combining practices in an HR system and conceptualizing and measuring the HR system at
different levels. Below, on the basis of our review, we highlight problems in current empirical
studies related to both of these areas, and for both, we offer a framework aimed to aid scholars
in refining theory and matching conceptualization and measurement.

80
Conclusion
We reviewed the empirical research on HR systems to date to identify trends and
progress over time and to pinpoint areas where progress is lacking. We used the
findings to identify directions for future research aimed toward further understanding
of how interrelationships between practices in an HR system affect outcomes
(summarized in Table 6). Most research to date does not align with the fundamental
assumption of synergies between HR practices in a system. The problems our review
highlighted in conceptualization at different levels, measurement, and combining
practices into systems hamper progress of the field in terms of understanding the
“system” element of HR systems. We offered two frameworks aimed at enhancing
conceptual clarity and construct refinement. The suggestions for future research from
these frameworks can help to develop less ambiguous and more rigorously developed
measures and build more specific theory and evidence on how practices interact within
HR

References
[1] Ahmad, S., & Schroeder, R. G. 2003. The impact of human resource
management practices on operational performance: Recognizing country and
industry differences. Journal of Operations Management, 21: 19-43.
[2] Andreeva, T., Vanhala, M., Sergeeva, A., Ritala, P., & Kianto, A. 2017. When
the fit between HR practices backfires: Exploring the interaction effects
between rewards for and appraisal of knowledge behaviours on innovation.
Human Resource Management Journal, 27: 209-227.
[3] Ang, S. H., Bartram, T., McNeil, N., Leggat, S. G., & Stanton, P. 2013. The
effects of high-performance work systems on hospital employees’ work
attitudes and intention to leave: A multi-level and occupational group analysis.
International Journal of Human Resource Management, 24: 3086-3114.
[4] Arthur, J. B. 1992. The link between business strategy and industrial relations
systems in American steel minimills. Industrial and Labor Relations Review,
45: 488-506.
[5] Arthur, J. B. 1994. Effects of human resource systems on manufacturing
performance and turnover. Academy of Management Journal, 37: 670-687.
[6] Arthur, J. B. 2011. Do HR system characteristics affect the frequency of
interpersonal deviance in organizations? The role of team autonomy and
internal labor market practices. Industrial Relations, 50: 30-56.
[7] Arthur, J. B., & Boyles, T. 2007. Validating the human resource system
structure: A levels-based strategic HRM approach. Human Resource
Management Review, 17: 77-92.
[8] Bae, J., & Yu, G.-C. 2005. HRM configurations in Korean venture firms:
Resource availability, institutional force and strategic choice perspectives.
International Journal of Human Resource Management, 16: 1759-1782.
[9] Becker, B., & Gerhart, B. 1996. The impact of human resource management
on organizational performance: Progress and prospects. Academy of
Management Journal, 39: 779-801.

81
[10] Becker, B., & Huselid, M. A. 1998. High performance work systems
and firm performance: A synthesis of research and managerial implications.
Research in Personnel and Human Resource Management, 16: 53-101.
[11] Becker, B., Huselid, M. A., Pickus, P. S., & Spratt, M. F. 1997. HR as
a source of shareholder value: Research and recommendations. Human
Resource Management, 36: 39-47.
[12] Birdi, K., Clegg, C., Patterson, M., Robinson, A., Stride, C. B., Wall, T.
D., & Wood, S. J. 2008. The impact of human resource and operational
management practices on company productivity: A longitudinal study.
Personnel Psychology, 61: 467-501.
[13] Böckerman, P., Bryson, A., & Ilmakunnas, P. 2012. Does high
involvement management improve worker wellbeing? Journal of Economic
Behavior & Organization, 84: 660-680.
[14] Böckerman, P., Bryson, A., & Ilmakunnas, P. 2013. Does high
involvement management lead to higher pay? Journal of the Royal Statistical
Society: Series A, 176: 861-885.
[15] Bonet, R. 2014. High-involvement work practices and the opportunities
for promotion in the organization. Industrial Relations, 53: 295-324.
[16] Boselie, P., Dietz, G., & Boon, C. 2005. Commonalities and
contradictions in HRM and performance research. Human Resource
Management Journal, 15: 67-94.
[17] Bowen, D. E., & Ostroff, C. 2004. Understanding HRM-firm
performance linkages: The role of the “strength” of the HRM system.
Academy of Management Review, 29: 203-221.
[18] Brown, A., Forde, C., Spencer, D., & Charlwood, A. 2008. Changes in
HRM and job satisfaction, 1998–2004: Evidence from the Workplace
Employment Relations Survey. Human Resource Management Journal, 18:
237-256.
[19] Bryson, A., Forth, J., & Kirby, S. 2005. High-involvement management
practices, trade union representation and workplace performance in Britain.
Scottish Journal of Political Economy, 52: 451-491.
[20] Camelo, C., Martín, F., Romero, P. M., & Valle, R. 2004. Human
resources management in Spain: Is it possible to speak of a typical model?
International Journal of Human Resource Management, 15: 935-958.

82
 Artificial Intelligence in Human Resource Management - Challenges and
Future Research Recommendations
Mandeep Kaur1
Franco Gandolfi2

Abstract
Digital innovation continues to fuel business transformation. Organizations have
realigned their strategic direction on enhanced adoption of digital technologies to leverage the
opportunities provided by the new age technologies, especially Artificial Intelligence. Enterprises can
leverage strategic advantage in talent - a key differentiator- by adopting Artificial Intelligence in
Human Resource Management. More than ever now, Human Resources function now regarded as a
trusted advisor, helping the organizations get through the transformational phase created by disruptive
technologies. This research provides insights on how Human Resources function has evolved as a
strategic partner by deployment of AI related technological advancements related, as it contributes to
building organizational capabilities and making organization competitive, thus creating organizations
that win in the market. It also looks at the challenges faced in Human Resource Management by
deployment of Artificial Intelligence. Insights are shared on future directions of potential research that
can be conducted in this field.

Keywords
Digital Transformation, Artificial Intelligence, Artificial Intelligence in HRM, Strategic HRM

JEL Codes: M12, D83, O33

DOI: 10.24818/RMCI.2023.3.382

1. Introduction
The advent of innovative and emerging technologies has shifted the boundaries of work
performed by humans, machines, and algorithms. The new age technologies have disrupted the
core business model in every industry. Digital transformation has resulted in enterprises
relooking at the very bedrock of who and what they are (Saarikko et al., 2020). All facets of
business and life have been affected by Artificial Intelligence (AI) and its associated
technologies (Heric, 2018). Artificial Intelligence may be defined, in its most basic forms, as a
machine's capability to learn from past experiences and adapt to fresh inputs and carry out tasks
much like humans. Since its inception, artificial intelligence (AI) has gone through ups and
downs (Kaur et al., 2022). AI improves problem-solving by translating data-

related inputs into electronic formats, including text, graphics, and numbers (Newell & Simon,
1972). Its computational powers, genetic algorithms, power, and data have been lauded as being

1 Dr. Mandeep Kaur, Director-HR, Kyndryl Corporation, Toronto, Canada, Email:

[email protected]
2 Dr. Franco Gandolfi, Corresponding Author, Distinguished Professor at California Institute of Advanced Management

(CIAM) and Georgetown University, Washington DC, USA, Email: [email protected]

Review of International Comparative Management Volume 24, Issue 3,

83
noteworthy for AI. (Yang, 2018). Artificial Intelligence plays a key role by being a linking
pivot to the cognitive technologies of mobile, robotics, and the Internet of Things (IoT) (Park,
2018). With AI, organizations can leverage competitive advantage by enhanced utilization of
their current human capital (Baker, 2020). AI has gained eminence in recent decades, the
technology advanced to the extent that there are AI-based systems that host AI decision-making
applications similar to human intelligence and are programmed with the ability to make
decisions (Galbusera et al., 2019). In the current context, Artificial intelligence (AI) can be
attributed to providing solutions to numerous challenges organizations face (Bharodiya &
Gonsai, 2018). Artificial intelligence is positioned as a cornerstone of work across industries,
as the past few years have witnessed an acceleration in its adoption. (Leopold et al., 2018). AI
now represents a field of research wherein research publications have grown exponentially. AI-
related conferences have multiplied, particularly in the previous ten years. (Kaur et al., 2021).
It is forecasted that by 2025, there will be an enhancement of proficiencies deployed of
algorithms and machines compared to the previous years (Leopold et al., 2018).
The transformation of HR from a support function to business partner in strategy has been
greatly aided by technology. in a way, redesigning people management (Rogers, 2018). The
adoption of AI is a necessity for organizations targeting an equitable workforce and is an
essential approach to address inconsistent conditions for HR function (Abdeldayem &
Aldulaimi, 2020).

2. Artificial Intelligence in Human Resource Management

Human Resource Management (HRM) has undergone a considerable transformation impacted

by technological changes (Bondarouk et al., 2017; Connelly et al., 2020). There has been a
repositioning of Human Resource Management as being more data-driven because of the
adoption of algorithmic technologies (Cheng & Hackett, 2021). Due to increased globalization
and technical advancements, the economic environment has witnessed significant changes.
Businesses today recognize that innovative and creative employees with organisational skills
provide an eternal competitive edge since, unlike other assets, their intellectual capital is
difficult for competitors to emulate. As a result, Human Resource Management now plays an
essential significant position. as it has a key role to play to attract, nurture, and engage talent.
As a result of the increased adoption of technology, the HRM function has become a crucial
partner. HR is at the forefront of technological changes, as digital technologies have impacted
workplace disruption. As a function, HR has needed to reimagine how work needs to be done
differently with Artificial Intelligence (AI) and the related cognitive technologies (Manuti &
Palma, 2018).
Human Resource Management (HRM) has undergone a
major transformation impacted by these large technological changes (Bondarouk et al.,
2017). As a function, Human Resources (HR) has needed to reimagine how work needs to be
done differently with Artificial Intelligence (AI) and the related cognitive technologies of
mobile, robotics, and IoT. (Manuti & Palma, 2018). AIT deployment in HR has contributed to
enabling organizations to gain and maintain a competitive advantage in the marketplace with
the right talent. AIT has brought about advanced functionalities to each facet of the Human
Resources function, resulting in a shift in how talent is managed in an enterprise (Bersin, 2018).

84
As algorithmic technologies are adopted, HR management is changing and evolving into being
more data-driven (Cheng & Hackett, 2021). AI technologies present numerous possibilities to
enhance HR functions. Integrating AI into HR functions enhances employee experience by
providing more capability, insights, and accurate information for decisive people management.
Analytical AI solutions assist with recruiting, selection, learning, growth, and retention in the
human resource function. (Kaur et al., 2021). Thus, as a priority, organizations need to focus
on value creation and be freed from transactional tasks (Kiron & Spindel, 2019). It has
introduced advanced functionalities to Human Resource Management in its entirety (Dhamija
& Bag, 2020). In their book section “Artificial intelligence techniques in human resource
management—a conceptual exploration”, Strohmeier & Piazza (2015) have emphasized on the
prospects in relation to AI in the Human Resource Management domain, enabled by the
advancement of specific functionalities of HR processes related to the prediction of turnover,
the search of potential candidates, scheduling of staff, conducting of sentiment analysis and
employee self-service. Walford-Wright & Scott Jackson (2018) highlighted the opportunities
created in the field of Talent Acquisition, due to advancement of technology. An industry report
by Ernst & Young (2018) highlights the importance of incorporating conversational AI in HR
transactions to perform administrative duties to build efficiency of the HR departments. HR
transactions can be automated and integrated by using cutting-edge technology securely. AI
can have a beneficial impact on the engagement and effectiveness of employees, as shared by
Smith (2019). On how AI improves efficiency in HR, Nicastro (2020), in his article on AI’s
integration into HR, highlights that AI is helping HR to reinvent AI practices to build
organizational efficiency, as AI applications support HR to analyze, predict and diagnose and
make better decisions. Yabanci (2019) has highlighted the concept of “Intelligent HR,” which
emerged because of the application of Artificial Intelligence technologies. It has impacted all
facets of Human Resources function. There is an extensive range of AIT applications in HR
function today, varying from chatbots for query resolution to predictive attrition analysis for
retention. Literature suggests that deploying AI in HR brings both immediate and long-term
implications. Insights about four aspects that AI contributes to, enable organizations to reduce
transactional work, recruiting, minimizing bias, and enhancing retention & talent mobility
(Sivathanu & Pillai, 2018). AI creates compelling employee experience that is a mirror of the
customer experience of an organization. (Meister, 2019). AI technology supports advanced
Sentiment Analysis which enhances understanding and enables insights into what employees
think and feel. Application of Neuro-Linguistic Programming (NLP) and Machine Learning to
open ended and text-based surveys, provides detailed insights to the HR leaders and managers
on their employees’ thinking and feelings (Strohmeier & Piazza, 2015). This transformation of
the Human Resources function has contributed to strengthening it to be recognized as a strategic
partner of the business in realizing organizational goals. Thus, Human Resources as a function
has evolved to become more digital, strategic, and innovative. Yet, as is true for any
transformation, it has been a long journey and has its challenges.

3. Challenges of Artificial Intelligence in Human Resource Management

The deployment of AI in HRM functions brings forth its own challenges. Literature suggests
that there was always a fear of automation affecting the adoption of technologies in workplaces
(Spencer, 2018). Although there is a growing influx of new technologies in HR, practitioners

85
have voiced caution in adopting it (Mathis, 2018). The field is still nascent and lacks theoretical
foundations and a clearly defined paradigm (Zhou et al., 2020). The literature currently
available on the adoption of AI with regards to HR is devoid of any solid conceptual
foundations (Prikshat et al., 2023).
The ethical debates surrounding AI technologies centre on a number of issues. One aspect of it
which contributes a significant portion of the opinions related to Artificial Intelligence revolves
around job loss. Each transformative new technology alters the market's need for specific job
roles. Organizations need to support employees in this transition to AI technology, by
recognising employee angst and reluctance and assisting them in dealing with it. Transparency
around jobs which will be impacted by AI, the components of jobs that are getting reduced and
those getting enhanced should be ensured by organisations.
Driving digital literacy and learning agility in employees needs to be prioritized to manage the
continuous reconstruction of work. (Park, 2018). Highlighting on four aspects of challenges of
AI in HR, first being intricacy of HR phenomena, second constraints brought about by limited
data sets, third regarding question of accountability linked to being fair and constraints
regarding ethics and legality, and fourth of possible negative employee reactions to decision
taken by management basis the algorithms based on data. (Tambe et al., 2019).
Deployment of AI initiatives has to put up with strong cultural and organizational barriers.
Leaders who are able to at the outset ensure that they break these barriers can effectively
leverage AI’s opportunities (Fountaine et al., 2019). This calls out another challenge that needs
to be handled and equip business leaders/HR professional which is related to understanding
that AI is not in competition but an enabler. Also, the need to build skills that can leverage what
is provided by AI as an enabler, is an issue that requires attention. These skills are related to
complex problem solving, critical thinking, creativity, and empathy.
While using technology in HR functions, concerns related to data security and privacy are
aspects that need to be looked at in organizations (Zafar, 2013). Data protection, security and
privacy concerns need to be addressed. These concerns could impact the flow of essential
operations impacting the organization (Yabanci, 2019). HR data privacy is one key aspect in
AI deployment in HR. Awareness and sensitivity to protection of data related to employees is
critical and organizations need to have governance guidelines to be set up. Such guidelines
should encompass the technical, data inputting processes and related legal and ethical issues.
(Ernst & Young, 2018). Many ethical concerns relating to the usage of artificial intelligence
has been highlighted by incidences of discrimination and bias across a spectrum of intelligent
systems. Although most firms strive towards automation with the best of intentions, integrating
AI into their hiring process may have some unintended repercussions. Black et al., (2021)
shared in their study that the adoption of AIenabled recruitment tools could lead to a rivalry for
these tools and the management consequences of these dynamics. There is need to address
challenges about the use of AI in the hiring procedures, such as what information should be
available when assessing an applicant for a position. AI systems in HR need to be designed in
a way that they address ethical concerns and ensure that the applications are aligned to the
broader societal value to be upheld and goals.
The deployment of AI is not sufficiently regulated, hence there is no enforcement mechanism
to ensure that it is deployed ethically. The negative consequences an unethical Artificial

86
intelligence system could have on corporations' financial lines are the current reasons for them
to adhere to these principles. Thus, resulting in a partnership between ethicists and researchers,
to have ethical frameworks developed to address this vacuum and manage the development and
societal adoption of AI models. Burton et al., (2017) recommended development of a
curriculum to introduce AI in academic institutions, which also included focus on ethics in AI.
AI technology when designed carefully and deployed with responsibility has potential to help
reduce bias. Building component of being fair into AI systems and complete transparency in
deployment of AI, will ensure that we remain on the on the positive side of the technology
spectrum separating good from bad (Guenole & Feinzig, 2019). Reimagining business process
with AI requires empowering workers with "fusion skills" so they can function well at the
humanmachine interaction. To leverage this collaboration, organizations have to comprehend
how humans can efficiently augment machines and to facilitate this collaboration, re-evaluate
business practises (Wilson & Daughtery, 2018). AI must not be perceived as a “cookie-cutter
solution” for all HR functions. (HRPA, 2017). AI augments, human decision making, does not
replace it. AI tools aim to extend human capabilities rather than automating or just replacing
them with machines (Jarrahi, 2018). Artificial intelligence (AI) tools do not replace humans in
decisionmaking. To assess distinctive characteristics, human intervention will always be
required to steer clear of pitfall of focusing on metrics rather than individuals. Focus on
upskilling of the HR teams, having ethical operating guidelines, and good technical curiosity,
the HR function can leverage all these, to effectively promote strategic advantage while
providing better experience to the employees, who put the strategy into action (Guenole &
Feinzig, 2019). Organizations need transform in case they need to completely leverage the
potential of AI (Wilson et al., 2017). In the long run, there is unlikely to be a substitute for this
transition (Hirsch, 2018). In this transformation, focus is by intelligent algorithms and
advanced, leveraging data of connected as well as smart products (Lichtenthaler, 2020). As is
true for any emerging technology, time and focus is needed to ensure that issues are addressed,
and benefits are maximized. Böhmer & Schinnenburg (2023) in their research critically
examined AI-driven HRM and organised context-specific capability building potential across
four inconsistencies that HRM must strategically solve to support the competitive edge of an
organisation, namely design of jobs, performance, transparency, and ambiguity of data. AI’s
capabilities are established and will continue to increase in HR, with time. The key is to ensure
that we not forget the human element of the role. Ensuring an open organizational culture is
also required, as implementation challenges need to be addressed, for organizations to integrate
the new technologies (Sivathanu & Pillai,2018). HR is going through massive disruption.
Employee experience bar has been redefined, due to impact of mobile and social technologies
and personalization. AI and automation are accessible to HR professionals and enable them to
solve several permeating talent concerns related to skills, retention, addressal of employee
concerns, objective job matching for internal and potential hires for career opportunities,
enabling managers with guidance on compensation investment, removing administrative tasks
through the robotic process automation, and facilitating an appealing platform for all the
employees to have flexibility to learn while being mobile (Guenole & Feinzig, 2019).

4. Future Research Recommendations

87
There has been a steady development of applications related to AI technologies for the HRM
function. Though there is limited research in some areas of this field, despite continual progress
being made in this field. Many areas merit additional investigation. By enabling digital
engagement, HR provides a competitive advantage to the organizations (Jesuthasan, 2017).
Cognitive enabled insights facilitate HR function to draw optimal outcomes by organizing work
and workforce in organizations. There is a dearth of research and detailed studies on this aspect
to validate this enablement of strategic HRM function, which contributes to supporting business
in optimal operational and strategic decisions.
There is a repositioning of HR Management, with its enhanced data-driven approach resultant
of adoption of applications based on AI and related algorithmic techniques (Cheng & Hackett,
2021). Further, detailed research work is recommended to be conducted on the model of
implementation of change for adoption of AI in HRM. Budhwar et al., (2022) recommended a
framework drawn from unified research HRM related AI applications and provided insights to
set a future research agenda.
Research is required on how AI and its related technologies of the HRM function have impacted
vital aspects of Employee Engagement, Retention, Growth, Compensation, Reward, and
Recognition. There have been very few studies conducted related to these aspects. Another area
which requires further exploration is on issues with implementing AI in HRM. It includes
studies on limitations of data mining due to small-size HR data sets, debates on ethics, fairness,
and possible adverse reaction of the employees to decisions based on data-based AI algorithms
related to people (Tambe et al., 2019). Aspects regarding the authenticity of employee data –
both current and potential, are a cause of concern as its validity is questionable. The authenticity
of algorithms designed based on the data could be imperfect, as it could reflect society's
ingrained prejudices and biases. Data sets could be structured in advance to be aligned to
historical precedents and patterns, which could even be part of an organization's culture and are
hardwired into code (Gulliford & Dixon, 2019). Questions regarding talent decisions made
basis this data, whether it further strengthens exclusions and existing biases, is imperative to be
researched, as these are sensitive topics to be addressed. Also, interlinked to this, there is a
requirement for more comprehensive insights and counsel in the form of additional research to
help address ethical concerns and acceptance of talent decisions based on AI applications.
There have been some of the drawbacks of data science being applied in HR highlighted,
including concerns related to infringing on privacy, usage of social media posts as a determinant
factor for hiring, which may lead to discriminative impacts on minorities/diversity.
Democratizing data, transparency, and providing data and insights to employees is another
aspect that needs further exploration (Hirsch, 2019). The skills of present and future HR
practitioners will need to be developed to manage today's AI applications and future
advancements. HR practitioners need to learn how to use AI-enabled analytical tools. Still, they
also need to be able to interpret and take action basis the analysis, thus developing numerical
analysis and reasoning skills to advise business (Davenport, 2019). HR professionals need to
have the competence to utilize technology to provide insights that support business success,
which necessitates the skill development of HR professionals (Wang & Lin, 2018). There are
hardly any studies conducted on this key aspect of AI in HRM, which is the new skills and
competencies that HR professionals need to be proficient in adopting and applying AI
applications in HRM and leveraging all the benefits. Chowdhury et al., (2023) emphasised that
to maximise the benefits of adopting AI in HRM, organisations must place a high priority on

88
the development of human competencies, leadership, an innovative work environment,
teamwork, and methods for integrating AI into the workforce.
Another avenue for future research is industry-specific and cross-industry comparisons to
support further research and insights related to the adoption of AI in HRM. There is a research
deficit as well related to the actual impact studies on the adoption of AI, Big Data, and related
technologies on the transformation of HR. Jatoba et al. (2019) has discussed AI applied to
Human Resources and distributed user behaviour. Research is required to delve into the aspects
of this HR transformation, thus providing insights, which can be potentially leveraged for future
growth and advancement of the HR function. How this phase of HR transformation and the
strategic development of HR has impacted business performance is a key area that has not been
much investigated. Future studies could examine how AI application in HRM affects
organisational abilities such Organisational Effectiveness, Knowledge Management, and
Innovative Performance.
Other predominant areas of research highlighted are Neural Networks, Fuzzy Logic, Decision
Trees, Classifiers and Evaluation Models. Research could be based on key aspects to feasibility
of leveraging of fuzzy logic in the Human Resource Performance Evaluation and for Talent
Acquisition and Selection, with fuzzy scripts being authored for Talent Evaluation for different
job roles. Another potential field of study is Sentiment Analysis and related aspect of the
leveraging of Text Analysis in the HRM context, wherein Sentiment Analysis is used as a
mechanism to decipher aspects that could not be detected by deploying standard objective
evaluation. Additionally, use of multiple related Computational Intelligence strategies which
are related to Data Mining techniques and Decision Trees is a vast area of potential research
which has been relatively untapped as of now. Also, while there has been some research
indicated on Individual Computational Intelligence Techniques, there is a lack of research on
combining of these techniques to leverage it as a cohesive model with all the techniques being
leveraged to their full advantage needs to be researched into. There is also a vast potential of
research to understand the impact of adoption of AIT, in all domains of the HR function, be it
Compensation/Payroll, Engagement, HR Services, HR support and transaction handing, as
adoption of AIT in HR is not consistent across all domains of HR. Future research in this area
may provide insight into the effects of implementing AIT in the HR function, on the key
organizational aspects related to Employee Engagement, Manager Enablement, and
Organizational Performance. Concerns pertaining to privacy and security of employee sensitive
information also needs deeper exploration. Challenges that HR professionals face, focuses on
security and privacy aspect. Threat posed either on confidentiality, integrity, and accessibility
front, in human-computer interaction, which impacts the HR professionals as it raises concerns
related to security, privacy, and related technological, legal, or ethical issues for individuals
and organisations. Technology plays key role in our lives today with self-driven cars or
robotics, but in HR it needs to be ensured that humans are not replaced, while technology grows
in the function. Emphasis that what HR professionals bring to the function, with empathy or
intuition and will taking decisions good judgment will always be key. Raising an important
question, that needs to be researched is that - with all the large technological advancements,
how can the HR professionals ensure that the pace of digital transformation, is at the right pace
for their business, so that very important human touch is not lost.

89
5. Concluding Thoughts

In the current environment, where the trends of globalisation and marketization, are continuing
to intensify, it is crucial to consider how to take advantage of the opportunity and gain a
competitive edge. Due to socioeconomic advancements, key talent has taken centre stage in
business development, and as a result, businesses are now vying for top candidates rather than
just the more limited pool of labour resources. Cognitive based intelligent systems like
Artificial Intelligence and related technologies support HR in achieving this. This integration
has opened by a large opportunity for the HR function to leverage computational intelligence
techniques such as Artificial Neural Network, Sentiment Analysis, Decision Tree Algorithms,
Fuzzy Logic to base talent decisions on. Notwithstanding all the benefits, AI is not flawless.
Nevertheless, there is requirement of human programming, therefore there is a possibility of
inaccuracies and biases. AI is indeed very good at analysing data and providing intelligent
recommendations for decisions, but it still leaves out important, non-technical components.
Despite the challenges of AI, a significant number of organisations are keen to integrate AI in
their Human Resources function because of the numerous advantages AI offers, which
outweigh its difficulties. Sakka et al., (2022) called out that AI for improving performance of
organisations can be successful when carefully implemented, with the right personnel hired,
HR employees retrained, and with a culture of inner transparency to prevent AI from being
used as a tool of control. Talent decisions are based on a greater degree of intelligence owing
to cognitive intelligence with the use of Artificial Intelligence technology, thus it offers a
promising strategy for Human Resource Management. Charlwood & Guenole (2022) in their
research called out that despite the emphasis on biases and inequity in popularized portrayals
of AI, these issues are easily remedied, emphasising that the to ensure that ethics and equity
stay at the forefront of AI for HR, HR professionals need to learn the necessary skills. The
challenges related to the adoption of AI due to budget constraints are being addressed by
development of enhanced intelligent and user-friendly interfaces, Machine Learning
algorithms, Pattern recognition, which support AI be within all businesses' budgetary limits.
Also concerted efforts are being made to make the system more secure, to address the
challenges and issues related to data privacy and security that HR has to manage with the
Artificial Intelligence applications. Key AI applications enable HR professionals to move away
from handling transactions/tasks which can be high volume and tedious, to focus on the
strategic aspect of Human Resource Management, thus enabling positioning of HR as a
profession/role to be strategic in nature. Especially in view of evolving nature of workplace
with hybrid model and higher focus on Inclusion and Diversity, the role of strategic aspect of
HR becomes more critical, which needs to leverage functionalities of AI in HR. Thus, support
organisations to be the “employer of choice of employees”- who have the skills - “that will
delight customers of the business”.
HR is no longer about people decisions being based on experience, intuitive or gut feeling.
Instead of only being process oriented, focus is now on being analytics driven. HR needs to
create a connected, transparent, mobile, personalized and 24x7 available universe, through the
workplace. Investing in AI and reinventing all the HR processes through technology and
looking at it from the lens of the employees, is required. In this digital era, HR must be
characterized by speed, personalization, and democratization.

90
References

1. Abdeldayem, M. M., & Aldulaimi, S. H. (2020). Trends and opportunities of artificial

intelligence in human resource management: Aspirations for public sector in Bahrain.
International Journal of Scientific and Technology Research, 9(1), 3867-3871.
2. Baker, N. (2020). Employee feedback technologies in the human performance system.
Human Resource Development International, 13(4), 477-485.
3. Bersin, J. (2018), “Talent trends technology disruptions for 2018”, Retrieved from
https://www.
isaconnection.org/assets/documents/2018BersinHRTechDisruptionsReport.pdf.
4. Bharodiya, A. K., & Gonsai, A. M. (2018). Research Review on Feature Extraction
Methods of Human Beings X-Ray Image Analysis. National Journal of System and
Information Technology, 11(1), 9.
5. Budhwar, P., Malik, A., De Silva, M. T., & Thevisuthan, P. (2022). Artificial intelligence–
challenges and opportunities for international HRM: a review and research agenda. The
International Journal of Human Resource Management, 33(6), 1065-1097.
6. Böhmer, N., & Schinnenburg, H. (2023). Critical exploration of AI-driven HRM to build
up organizational capabilities. Employee Relations: The International Journal.
7. Bondarouk, T, Parry, E, & Furtmueller, E. (2017). Electronic HRM: four decades of
research on adoption and consequences. International Journal of Human Resource
Management, 28 (1), 98-131.
8. Black, J. Stewart, and Patrick van Esch. (2021) "AI-enabled recruiting in the war for
talent." Business Horizons,64 (4), 513-524.
9. Burton, E., Goldsmith, J., and Koenig, S. (2017). Ethical considerations in artificial
intelligence courses. AI Magazine, 38, 22–34. doi: 10.1609/aimag.v38i2.2731.
10. Charlwood, A., & Guenole, N. (2022). Can HR adapt to the paradoxes of artificial
intelligence? Human Resource Management Journal, 32(4), 729-742.
11. Cheng, M. M., & Hackett, R. D. (2021). A critical review of algorithms in HRM:
Definition, theory, and practice. Human Resource Management Review, 31(1), 100698.
12. Chowdhury, S., Dey, P., Joel-Edgar, S., Bhattacharya, S., Rodriguez-Espindola, O.,
Abadie, A., & Truong, L. (2023). Unlocking the value of artificial intelligence in human
resource management through AI capability framework. Human Resource Management
Review, 33(1), 100899.
13. Connelly, C. E., Fieseler, C., Černe, M., Giessner, S. R., & Wong, S. I. (2020). Working
in the digitized economy: HRM theory & practice. Human Resource Management Review,
31(1), 100762.
14. Davenport, T. H. (2019). Is HR the most analytics-driven function? Harvard Business
Review. Accessed August 10, 2020 from https://hbr.org/2019/04/is-hr-themostanalytics-
driven-function.
15. Ernst & Young (2018) - The new age: artificial intelligence for human resource
opportunities and functions. Accessed August 5,2020 https://www.ey.com/Publication/
vwLUAssets/EY-the-new-age-artificial-intelligence-for-human-resourceopportunities-
and-functions/$FILE/EY-the-new-age-artificial-intelligence-for-humanresource-
opportunities-and-functions.pdf.
16. Fountaine, T., McCarthy, B., & Saleh, T. (2019). Building the AI-powered organization.
Harvard Business Review, 97(4), 62-73.

91
17. Galbusera, F., Casaroli, G., & Bassani, T. (2019). Artificial intelligence and machine
learning in spine research. JOR spine, 2(1), e1044.
18. Guenole, N., & Feinzig, S. (2019). The Business Case for AI in HR with Insights and Tips
on Getting Started. IBM Smarter Workforce Institute.
19. Gulliford, F., & Dixon, A., (2019). AI: the HR revolution. Strategic HR Review,
https://doi.org/10.1108/ SHR-12-2018-0104.
20. Hirsch, P.B. (2019), Tinker, Tailor, Soldier, Spy, Journal of Business Strategy, 40 (4)53-
56.
21. HRPA. (2017). A new age of opportunities: what does artificial intelligence mean for HR
professionals? Accessed August 2, 2020, from
https://www.hrpa.ca/Documents/Public/Thought-Leadership/HRPA-Report-
ArtificialIntelligence-20171031.PDF.

92
Scientific Programming 100

Hindawi
Scientific Programming
Volume 2021, Article ID 8387277, 10 pages https://doi.org/10.1155/2021/8387277

Research Article
Research on Human Resource Recommendation Algorithm
Based on Machine Learning

Hong Zhu 1,2 1 Xi’an Jiaotong

University, Xi’an, Shaanxi 710049, China 2

Shaanxi XueQian Normal University, Xi’an, Shaanxi 710100, China

Correspondence should be addressed to Hong Zhu; [email protected]

Received 21 June 2021; Revised 2 August 2021; Accepted 13 August 2021; Published 25 August 2021

Academic Editor: Muhammad Usman

Copyright © 2021 Hong Zhu. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original
work is properly cited.

The economic environment has changed dramatically around the world in recent years, generating favorable
conditions for the growth of small- and medium-sized firms. The socioeconomic development and international
integration of China are greatly influenced by the growth in both quality and quantity, the scale of operations,
and the internal force of small- and medium-sized businesses. Moreover, in comparison with other developed
countries around the world, Chinese small- and medium-sized enterprises continue to face many limitations in
terms of size and contribution levels and have not yet fully realized their potentials due to difficulties and poor
quality; human resources in this field are still lacking. This study defines the current state of human resources in
small and medium firms, the factors that impede development, and the steps that can be taken to overcome
these obstacles in order to assist human resource development in this sector during the current period. This
study uses machine learning (ML) techniques to manage and analyze human resource data in modern
enterprises. The ML techniques realize the functions of the human resource system and reduce the business
volume in human resource in order to improve the efficiency and management of the human resource work. In
this paper, we designed and implemented the wage forecasting model in human resources that uses a gradient
descent algorithm, its types, and backpropagation (BP) neural network to improve the accuracy of the
forecasting model. We performed multiple experiments by using a various number of neurons in the hidden
layers, different number of iterations, and several types of gradient descent algorithms. The BP neural network
model was performed brilliantly by attaining the training accuracy of 89.98% and validation accuracy of 84.05%.
The experimental results show the significance and

93
Scientific Programming 101

importance of the proposed work. task for human beings to implement at all, so the
existence of information management system (MIS)
came into being. Management information system
(MIS) is management software having the
Introduction functionality of data collection, database creation, and
Human resource management system (HRMS) is an data management (including data storage and
important tool of enterprise management that plays extraction) and can analyze a large amount of data [2–
an important role in the development of enterprises. 4]. The HRMS is an information management system
From its birth to now, HRMS has experienced three which combines both HRM and information
stages [1]. The first generation of HRMS appeared in technology (IT). It is not only an information
the 1960s. The main objective and functionality of processing tool but also a resource management
HRMS at that time was to automatically calculate the standard. Its main purpose is to standardize the
salaries of employees via a computer system, business process of the human resource department
excluding salary data analysis and nonfinancial through the system, concentrate human resource
information. The main reason for such circumstances information, and improve the transparency of HRM.
at that time was the limited resources and technical HRMS plays an important role in enterprises and has
requirements. The second generation of HRMS started the powerful functions of optimizing business
in the early 1980s. The main drawbacks of the first processes, improving work efficiency, and improving
generation were overcome in the second generation management mode. Its quality is directly related to
with the development of the database technology and the performance of the enterprise, which affects the
other requirements missing in the first generation. survival and development of the enterprise [5, 6].
The functions of nonfinancial information, salary data Therefore, the HRMS is considered an important tool
analysis, and report production were kept in for enterprises. However, in recent years, the data
consideration, which greatly improved the collected by enterprises are growing rapidly and
practicability of HRMS in the enterprises. However, continuously. Such a large amount of data has
there were still some major problems in the second exceeded the processing capacity of HRMS, which
generation of HRMS; for example, the management leads to its failure to conduct data management and
process was not standardized and the information data analysis normally. This situation is described as
processing techniques were not comprehensive. The “rich data, but the information is poor.” Furthermore,
third generation of HRMS appeared in the late 1990s. the majority of the enterprise-level HRMS has not
Because of the fierce competition among enterprises been fully utilized, with only data collection and
at that time, especially in the introduction and storage, no effective data management and analysis,
cultivation of talent, HRM has risen to the top of the and no use of a large amount of data, wasting not only
company’s priority list. Due to the popularity of the enterprise’s limited resources but also the
computers at that time, different technologies, opportunity to use big data to develop rapidly. Due to
particularly database technology and Internet, had the fact that the traditional HRMS cannot fulfill the
made great progress in the HRMS. These technologies needs of organizations’ while having big data, in order
consolidate human resources for unified management to make the system work normally, it must artificially
and create integrated data sources, making data screen and process such a vast amount of data,
analysis and sharing easier. thereby reducing the volume of big data. However,
this kind of processing method is back to the old era of
With the advent of the information age and the manual data processing, which seriously affects the
development of enterprises, a vast amount of data is improvement and efficiency of the work. Therefore,
produced which needs to be managed and controlled people started the use of data mining techniques to
in a well-organized manner. To control, manage, and reduce the data to be processed [7, 8]. The purpose of
analyze such a large amount of data is indeed a tough data mining is to analyze and find out the important

94
Scientific Programming 102

data patterns in the massive amount of raw data and Some of the basic contributions of this paper are given
finally convert the original data into useful information as follows:
and knowledge.
(1) This study designs and implements the model
The 21st century is the era of big data that brought a of human position matching using ML
revolution in every aspect of human life such as algorithms, which can provide a guarantee for
effective human resource allocation.
healthcare, industries, government sectors, social
media, and business organization and changed the (2) A new method is proposed to evaluate the
candidates in an online recruitment system and
policies of human resource management [9, 10]. The
to solve the problem of candidate ranking using
traditional HRMS can realize the functionality of data ML algorithms. Further, a new algorithm based
entry, query, and statistics but cannot analyze the on support vector regression is designed for the
relationship between data samples effectively and prediction of the human resource demand of
cannot predict the future development of enterprises enterprises which has a good practical
according to the existing data. To solve this problem, reference value.
people put forward using data mining techniques to (3) To solve the problem of staff turnover, an
extract and analyze big data. Machine learning (ML) is extreme gradient boosting (XGBoost) algorithm
one of the main tools to realize data mining, and it has is used to predict the employee turnover rate.
various applications in different fields, including the It is proved with the help of experimental
results that it is more accurate than the other
field of HRM and businesses. The important roles of
supervised classification algorithms in
ML in human resource management include predicting employees’ turnover.
recruitment, talent management, preventing brain
(4) Based on the work and life situation, this study
drain, and improving human resource precise analyzes the factors affecting the retirement of
management and are analyzed in detail by different employees and uses ML algorithms to create a
researchers [11, 12]. In a study conducted by Chen et retirement prediction model, which can
al. [13], they presented the capability of data mining provide solutions for the outflow of human
techniques in the development of the decision-making resources. Further, this study realizes a human
process of HRMS. The significance of data mining in resource ranking model, which can predict the
turnover management is analyzed and implemented ranking and classification of resumes with high
accuracy, and effectively simplifies the work of
by Meng [14]. They presented a knowledge discovery
human resources.
approach to assist firms in predicting the future
performance of their employees and then assigning The rest of the paper is organized as follows. Section 2
relevant personnel to the appropriate project represents the related work, while Section 3 illustrates
positions. the applications of the backpropagation neural
network (BPNN) in wage forecast. Section 4
In this paper, a fuzzy clustering method based on an demonstrates the experimental simulations and
adaptive genetic algorithm is proposed that provides results analysis, and finally, Section 5
effective data analysis for the performance appraisal
of employees in human resource management of concludes the research work.
modern enterprises. The potential of AI in HRM is
explored in much detail in different studies, among
1. Related Work
which six scenarios are turnover prediction of the
artificial neural network (ANN), candidate search The term “human resources” was initially coined in the
based on a knowledge search engine, personnel early 1970s, when the importance of labor relations
scheduling of genetic algorithm, HR emotion analysis, became more widely recognized, and concepts like
resume data collection and extraction of information, motivation, organizational behavior, and selection
and interactive voice response employee self-service. assessments began to emerge. In “The Human
Resources Glossary,” Huong [15] describes human

95
Scientific Programming 103

resource management as “the people and staff who behavior, as well as dramatic changes in the structure
administer an organization,” as opposed to an and intensity of worldwide competition between large
organization’s financial and material resources. ML is organizations, had a significant impact on personnel
a subject that studies how to improve the management methods near the end of the 20th
performance of the system itself using calculation and century, and as a result, HRM emerged [21].
experience. In general, it transforms the unordered
HRM has become the standard technique for
data into useful information through calculation, and
managing people inside a business since the early
this information generally refers to the model
1980s [22]. It has been highlighted as an important
obtained through training and learning. With the use
method for developing management systems that
of computer systems for human resource tasks, HRM
effectively steer human competencies toward
has experienced significant changes since the 1980s.
corporate goals since then [23]. With the rise of the
Initially, human resource professionals used these
Internet as a professional platform and technological
systems primarily as record-keeping systems for
advancements, human resource information is
simple operations like accepting and storing job
evolved into a central hub for human resource
applications on behalf of their employers [16].
management. Electronic HR is the more generic term
However, as information technology (IT) advanced, HR
for the use of computer technology in HR [24]. E-HRM
professionals began to employ sophisticated
appears to have become the prevalent word for
computer systems to do the majority of their everyday
referring to virtual HR or e-enabled HRM in the last
tasks and make strategic decisions. As a result, HR
two decades, in which IT is used to implement
procedures have been greatly simplified [17]. IT is
comprehensive HR functions within a business [25,
increasingly and widely used in HRM, and complex
26]. In general, it is regarded to be useful since it is
computer systems are required for both internal and
designed to improve HR efficiency, save costs,
external operations such as recruitment, performance
minimize administrative responsibilities, simplify HR
assessment, remuneration, outsourcing,
planning, and empower HR practitioners to become
compensation, onboarding, development, strategic
strategic or business partners in firms. According to
analytics, and business-to-employees (B2E). The
Marler and Parry [27], e-HRM may not be as cost-
advancement in computer engineering has resulted in
effective as it claims to be. Organizations have been
the emergence of advanced technologies like artificial
transitioning from traditional faceto-face HR
intelligence (AI). AI is also bringing up significant
procedures to current electronic ones; e-HRM is
changes. HRM, a tech-dependent phenomenon, has
becoming an important theoretical and practical
recently seen such a shift. These significant shifts in
phenomenon [28, 29]. Many individuals expected e-
HRM that we are witnessing are well worth
HRM to revolutionize the way HRM is done in
investigating. The industrial revolution in the 18th
organizations, moving it away from being purely
century caused a significant shift in industries and
administrative and toward being more strategic.
economics [18]. These significant shifts in HRM that
According to Thite et al. [30], the electronic transition
we are witnessing are well worth investigating [19].
faces two major hurdles in the perspectives of
Manufacturers encountered the scientific
business and technology leaders. One is HR’s ability to
management paradigm in the early 20th century,
assist business leaders in adopting a digital mindset, as
which promoted standard-performance production as
well as a digital way of managing, organizing, and
the optimum approach to achieve maximum
leading change. The second is HR’s ability to transform
profitability [20]. Employers at the time were more
the whole experience of employee by transforming HR
concerned with employee production rather than with
processes, systems, and the HR organization via novel
employee satisfaction, owing to the application of
digital platforms, applications, and way to deliver HR
scientific management techniques and the lack of
services.
governmental regulation and law governing personnel
policies and practices. The widespread study of human

96
Scientific Programming 104

Information technology (IT) is on many people’s the problem of human resource replacement in
agendas, especially professionals and academics, as a technical posts, this study uses supervised ML
crucial commercial tool for many firms. What is more techniques to predict the appropriate human
intriguing is that there’s a global effort underway to resources to fill the vacancy of the technical posts. All
improve IT’s qualities using AI, often known as of these indicate that the influence of ML in the field
cognitive computing. According to Pan [31], since of human resource management is expanding and has
2010, several IT behemoths have bought a great potential to improve the efficiency and
approximately 140 AI start-ups. These businessmen performance of HRMS.
are attempting to combine artificial intelligence (AI)
into a variety of computer systems in order to produce 2. Applications of ML and
seamless experiences during the installation of these
technologies. AI strives to make machines capable of BackpropagationNeural Network
performing tasks in the same way that humans do. (BPNN) in Wage Forecast
However, given the fact that people have not yet
According to R. Michalski, research in the field of ML
completely deciphered and grasped the human
can be divided into “learning from examples,”
intelligence that AI simulates, Shah and Warwick [32]
“learning in problem-solving and planning,” “learning
believe that computers functioning like human beings
through observation and discovery,” and “learning
should be viewed in terms of imitation. The Japanese
from instructions.” According to E. A. Feigenbaum, in
Cabinet has announced a new project called Society
the manual of AI, ML can be divided into “mechanical
5.0 with the goal of bringing about societal
learning,” “teaching-learning,” “analogy learning,”
improvement. This new approach appears to be aimed
and “inductive learning.” At present, inductive
at achieving a high level of human-machine interaction
learning is the most widely used learning technique of
[33]. Its goal is to connect people, objects, and
ML. There are three main streams of inductive
computers in cyberspace, using sensors to detect
learning: connectionist learning based on neural
commands and AI to execute them [34]. Unlike the
networks, semiotic learning, and statistician learning.
idea of Business 4.0, which we are approaching,
Connectionism learning based on neural networks is
Society 5.0 is not limited to the industrial industry but
learning of the ML model that simulates the
rather penetrates social life as AI, the Internet of
functionality of the neurons of the human brain, and
Things (IoT), augmented reality, and robotics come
its work in terms of ML is represented by the
together [35].
perceptron, although connectionism learning started
In this paper, a new HR recommendation algorithm early and encountered great obstacles. The main
based on the hidden semantic model and the deep reason was that the neural network at that time could
forest is proposed in document, which can better only deal with a simple linear problem and even could
recommend the interesting posts for the users. not solve the small problem of XOR, which led to the
Various studies made an attempt to understand that development and learning of the neural networks.
whether machine learning (ML) can be effectively Semiotic learning was originated from mathematical
used to perform potential assessments using logic, and its core idea is to predict the results through
preventive methods, thus helping or even replacing the deduction and inverse deduction of symbols,
human resource managers. The ML methods are used which mainly represents decision tree and logic-based
in the earlier studies to provide human resource learning. Statistical learning is used to apply statistical
decision support. In this paper, a model of employee learning knowledge to ML techniques, and its
turnover based on a gradient descent algorithm and representative technology is SVM and “kernel
BP neural network is established in document, which method.”
can predict the turnover of employees in enterprises
Neural network is a parallel interconnected network
and provide enterprises with the opportunity to solve
composed of adaptive simple units, and its
any problems and improve the retention rate. To solve
97
Scientific Programming 105

organization can simulate the interaction of biological applications. In general, the BPNN model refers to the
neural system to realworld objects. The “simple unit” multilayer feed-forward neural network trained by the
in the definition of a neural network refers to the BP algorithm, as shown in Figure 2.
neuron model, and the neuron model generally refers
to the “MP neuron model or single-layer perceptron
1.ModelSelection. Salary prediction is a kind of
model” proposed by McCulloch and others in 1943
regression problem. The regression prediction models
[14]. In the model, firstly, the neurons get input signals commonly used in ML include linear regression,
from the environment and then multiply them with polynomial regression, ridge regression, lasso
the corresponding weights put on the edges. The regression, elastic network regression, and neural
output of a neuron then becomes the input for networks. Linear regression is a regression model
another neuron and this process continues until it composed of linear variables. Therefore, the linear
reached the last layer. Therefore, neurons can relationship between the independent variables and
calculate the total input signal by multiplying itself dependent variable must be satisfied; that is, linear
regression is only applicable to solve linearly separable
with the corresponding weights, then compare it with
problems. In linear regression problems, there exists a
the threshold value of neurons, and process it via the linear relationship between variables; as a result, the
activation function. At last, the output of the neuron is calculation complexity is low and the training speed is
obtained, as shown in Figure 1. faster. The disadvantage of linear regression is that it
is very sensitive to outliers. If there is any outlier in the
A neural network is a topological structure consists of data, the accuracy of the model will be affected.
various neurons connected via a particular hierarchical Polynomial regression is the extension of linear
structure. Neural network models can be classified regression, which refers to the regression model with
into two types based on the hierarchical structure of independent variables, having an index greater than 1
neurons: interconnection network model and in the regression equation. When the independent
hierarchical network model. The neurons in the index is equal to 1, the polynomial regression equation
degenerates and becomes a linear regression
hierarchical network model are hierarchical and are
equation. As in most of the cases, independent
generally divided into several layers according to their variables and dependent variables are not linear, so
functions, and each layer is connected, such as the the polynomial regression is an automatic choice for
input layer, the middle layer, and the output layer. The it, which can solve nonlinear and complex problems
interconnection network model is that in which any with ease. The disadvantage of the polynomial
two neurons can be connected and has certain regression model is that the best index is difficult to
randomness. The hierarchical network has become determine, the high index is easy to cause overfitting,
the most widely used model structure because of its and the low index is prone to underfitting.
Ridge regression is an optimization technique used
good structure and easy analysis.
when the data features of linear regression or
As mentioned earlier, the learning process of the polynomial regression have multiple collinearity (i.e.,
neural network is to adjust the parameters’ values of there is an approximately linear correlation between
neurons, and the backpropagation (BP) algorithm is independent variables). It reduces the influence of
used to achieve this purpose in order to reach the multiple collinearities by adding a square bias factor to
target. BP algorithm is based on gradient descent the regression estimation function (loss function).
strategy, which takes the negative gradient direction Lasso regression is similar to ridge regression, and it is
of the target as the search direction to adjust the also an optimization technique used when the data
parameters and finally achieves convergence by features of linear regression or polynomial regression
continuously iterating and updating the parameters. It have multiple collinearities and also add bias factors to
is one of the most successful neural network regression estimation function. Only the square
algorithms and has been used widely in various

98
Scientific Programming 106

1.ModelSelection. Salary prediction is a kind of equation. As in most of the cases, independent

regression problem. The regression prediction models variables and dependent variables are not linear, so
commonly used in ML include linear regression, the polynomial regression is an automatic choice for
polynomial regression, ridge regression, lasso it, which can solve nonlinear and complex problems
regression, elastic network regression, and neural with ease. The disadvantage of the polynomial
networks. Linear regression is a regression model regression model is that the best index is difficult to
composed of linear variables. Therefore, the linear determine, the high index is easy to cause overfitting,
relationship between the independent variables and and the low index is prone to underfitting.
dependent variable must be satisfied; that is, linear Ridge regression is an optimization technique used
regression is only applicable to solve linearly separable when the data features of linear regression or
problems. In linear regression problems, there exists a polynomial regression have multiple collinearity (i.e.,
linear relationship between variables; as a result, the
there is an approximately linear correlation between
calculation complexity is low and the training speed is
independent variables). It reduces the influence of
faster. The disadvantage of linear regression is that it
is very sensitive to outliers. If there is any outlier in the multiple collinearities by adding a square bias factor to
data, the accuracy of the model will be affected. the regression estimation function (loss function).
Polynomial regression is the extension of linear Lasso regression is similar to ridge regression, and it is
regression, which refers to the regression model with also an optimization technique used when the data
independent variables, having an index greater than 1 features of linear regression or polynomial regression
in the regression equation. When the independent have multiple collinearities and also add bias factors to
index is equal to 1, the polynomial regression equation

Figure 1: Single-layer perceptron model.

Figure 2: General architecture of backpropagation neural network (BPNN) model.

degenerates and becomes a linear regression regression estimation function. Only the square offset

99
Scientific Programming 107

is added in ridge regression, and the absolute value confirmed according to the performance of the model.
offset is added in the lasso regression. Elastic net There are many empirical formulas used to calculate
regression is a mixture of ridge regression and lasso the number of neurons in the hidden layer. The most
regression. It adds both square bias and absolute value commonly used formulas to calculate the number of
offset in regression estimation function, to achieve the neurons in the hidden layer are given as follows:
common effect of ridge regression and lasso
√
regression. BP neural network is the most widely used
nh ni · no, (1)
type of‘ neural network. Theoretically, a three-layer
BP neural network can approach any continuous √ nh
function with arbitrary precision, has a strong learning ni · no +
ability and simple network structure, and takes low k, (2)
time in calculation. At the same time, neural networks where ni is the number of input layer nodes, no is the
have some shortcomings too. The most important and number of output layer nodes, k [1, 10]. According to
commonly faced problem in neural networks is the the empirical formula (2), the range of the number of
topological structure, parameters initialization, and neurons in the hidden layer is estimated to be [5, 14],
optimization techniques of neural network. Further, it then space is slightly expanded to [3, 16] during the
has a great impact on the training results of the model; training, and the BP neural network is used to train
if the selection of superparameters is not good, it will repeatedly.
significantly affect the performance of the model. To sum up, the number of neurons in the input layer,
hidden layer, and output layer is 14, 15, and 1, the
3.2. Salary Forecast Model Based on BP Neural
activation function of neurons is the sigmoid function,
Network. The BP neural network offers nonlinear
the error function (objective function) is a quadratic
mapping, generalization, and fault tolerance
mean square function, Gaussian random variable with
capabilities. Moreover, the multilayer (three-layer or
a mean value of 0 and standard deviation of nin is
above) BP neural network can approximate any
selected as initialization weight, and Gaussian random
precision nonlinear function theoretically, which is a
variable with the mean value of 0 and standard
powerful ML algorithm. To establish a salary forecast
deviation of 1 is selected as initialization threshold.
model based on BP neural network, it is necessary to
solve the key problems such as topology, activation
cj f′􏼐netj􏼐􏼐 δvjk lj􏼐1 − lj􏼐􏼐 δvjk, (3)
function selection, and parameter initialization.
k k where yj is the target value

The design of the network topology is relatively free of the K nodes and zk is the output value of the K
and has no fixed form. The hierarchical structure is nodes. After the salary forecast model is established,
generally designed according to the nature of the the training can be started. The training flowchart of
problem. In most cases, the network parameters are the model is represented in Figure 3.
also selected according to the nature of the problem In the previous introduction, BP neural network
and earlier experience. Salary prediction is a mapping adopts the gradient descent method in the reverse
of multiple input and a single output. According to the propagation update. According to the amount of data
format of data in demand analysis, it can be used in each update, the gradient descent method can
determined that the number of input layer neurons in be divided into three forms: batch gradient descent
the network is 14 and the number of neurons in the method, random gradient descent method, and small-
output layer is 1. There is no definite calculation batch gradient descent method. Each update is called
method for the number of neurons in the hidden batch gradient descent which uses all samples to
layers. Some methods can be used to estimate its update. Because batch gradient descent needs to use
approximate range, and then the model can be trained all training samples at every iteration, when the
repeatedly using the trial and error method. Finally, number of training datasets is large, the time of one
the number of neurons in a hidden layer can be iteration update will be longer, the cumulative error of
100
Scientific Programming 108

all samples will decrease down to a certain extent, and the simulation results. The list of python packages
eventually, the training process will be slow. As a used in this study was Keras, TensorFlow, seaborn,
result, the batch gradient descent training process Numpy, Matplotlib, and Sklearn, and so on.
may fall into local optimal but cannot achieve global
optimization. The method of updating only a single 4.1. Results of Salary Forecast Model Based on BP
sample is called stochastic gradient descent (SGD). Neural Network. This subsection demonstrates the
Because SGD only uses a single sample for each simulation results carried out via the salary forecast
iteration update and causes parameter updates very model based on BP neural network. All the simulation
frequently, the update between different sample data results were attained using a different number of
may offset each other, which makes each update not neurons and iterations as shown in Tables 1 and 2.
advance toward the overall optimization direction, Table 1 shows the training and validation accuracies of
and it is prone to oscillation phenomenon, but this the salary forecast model based on the BP neural
oscillation also makes it possible to jump out of local network, using the data size of 1000 samples. The
optimal position and reach global optimal position. dataset samples were divided into training and testing
The minibatch gradient descent (MBGD) is a samples, the size of the training dataset was 700, the
compromise method to solve the problems existing in size of the testing dataset was 300, the number of
BGD and SGD. It divides the whole sample dataset into neurons in each hidden layer was trained 30 times,
small sample datasets and takes the cumulative error and finally, the average value of the 30 times results
gradient of each small sample dataset as the was calculated. The execution process reached its
estimation value of the real gradient of the whole termination when the number of iterations reaches
sample. When the sample size is large enough, the 250.
local error gradient value is approximately equal to the
From Table 1, it is quite obvious that the training and
total sample error gradient value. The small-batch
validation accuracies increase as the number of
gradient descent method can have both bad and SGD
neurons in the hidden layer is increased. Figure 4
advantages.
shows the training and validation accuracies of the
salary forecast model based on the BP neural network.
3. Experimental Simulations
andResult Analysis Figure 4 illustrates that the training and validation
accuracies increase with the increase in the number of
This section of the paper represents the experimental iterations. The highest training accuracy of 89.98 was
results carried out via different experiments and their attained at the 250th iteration. The highest validation
analysis. All the experiments were performed on a accuracy of 84.05 was also conquered at the 250th
laptop computer system having the following iteration.
specifications: Intel Core-i5, 7th generation, processor
of 2.4 GHz, 16 GB of RAM, and the operating system Table 2 shows the training and validation loss of the
installed on the system being Microsoft Windows 10). salary forecast model based on the BP neural network.
Anaconda Jupyter Notebook is used as an IDE for The training and validation loss of the BP neural
carrying out the simulations. Python has been used as network decrease as the number of neurons in the
a language for the implementation and for generating hidden layer is increased.

101
Scientific Programming 109

Establishing model and initialization

Calculate the output of hidden layer and

output layer

Adjusting parameters of hidden layer

The number of

generation is Stop condition

error is less than

Figure 3: Training flowchart of the wage prediction model.

Table 1: Training and testing accuracies of BP neural network We also

Number of neurons Training accuracy Validation
accuracy
13 89.00 82.17
3 86.66 80.04
14 89.50 83.08
4 87.04 80.20
15 89.80 83.86
5 87.30 80.44
16 89.98 84.05
6 87.44 80.52
performed multiple experiments using different using a different
7 87.46 80.72
number of neurons in the hidden layer. gradient descent techniques,
8 87.80 80.96
9 87.94 81.22 that is, batch gradient descent (BGD), stochastic gradient
10 88.00 81.44 descent (SGD), and minibatch gradient descent (MBGD). To
11 88.52 81.73 show the training effect of BGD, SGD, and MBGD in an
12 88.64 81.93 effective way, the loss function curve of these techniques
102
Scientific Programming 110

during training is shown in Figure 5. Among them, the horizontal coordinate in Figure 5 is the iteration update
times, the ordinate is the loss function (cost), the training sample size is 200, and the small-batch sample size is
set to 20. So, BGD performed 200 cycles (read 200 rounds of dataset), MBGD performed 20 cycles, and SGD
performed a single cycle.

From Figure 6, it can be seen that when the iteration update time is the same, the convergence results of the
three gradient descent methods are relatively close, but the loss change curve is very different: SGD curve has a
large fluctuation, BGD curve is relatively smooth, and MBGD is
Table 2: Training and validation loss of BP neural network using a
different number of neurons in the hidden layer.
Number of neurons Training loss Validation loss
3 0.009975 0.010725
4 0.009958 0.010724
5 0.009950 0.010723
6 0.009954 0.010720
7 0.009947 0.010716
8 0.009944 0.010710
9 0.009940 0.010690
10 0.009938 0.010660
11 0.009932 0.010630
12 0.009928 0.010595 Iteration #
13 0.009926 0.010580
14 0.009920 0.010553
15 0.009916 0.010510
Figure 5: Training and validation loss of BP neural network.
16 0.009912 0.010502 Comparision
Training and Validation accuracy

0 25 50 75 100 125 150 175 200

Iteration update times
100 90 80 70 60 50 40 30 20 SGD
10
BGD
0
50 100 150 200 250 MBGD
Iteration #
Figure 6: Comparison of SGD, BGD, and MBGD iterative update
Training accuracy process.

Validation accuracy
between SGD and BGD and is relatively stable. The
Figure 4: Training and validation accuracies of BP neural network.
training speed of SGD, BGD, and MBGD is different,
and the training period of SGD, BGD, and MBGD is
different, which is different from that of SGD, BGD,
and MBGD. If the sample space is large, the training

103
Scientific Programming 111

speed of the investigated three techniques will be

more obvious, especially SGD and BGD. In short, the
training speed of SGD is fast and iteration update is
frequent, but there is a fluctuation that easily
produces oscillation. BGD is stable and iteration
update is less, but the training speed is slow. MBGD is
a compromise scheme, which has a fast training speed
and is relatively stable. 0 20 40 60 80 100 120 140
Epoch
The three-layer BP neural network model used for the
SGDRMSProp
salary prediction consists of a different number of MomentumAdam
neurons in different layers. The number of neurons in NAGNadam
Adagrad
the input layer, hidden layer, and output layer is 14,
15, and 1, the activation function is the sigmoid Figure 7: Comparison of convergence rates of various optimization
methods.
function, the error function is quadratic mean square
error (QMSE), and the network parameters are
updated by using small-batch gradient descent
method (SBGD) and mixed optimization method
Nadam. The method used in this experiment divides
the dataset into two parts, that is, training and testing.
The total sample space selected for the experiment
was 1000, of which 900 samples were selected for
training while the rest of the 100 samples were
selected for testing and validation purpose.

Figure 7 shows the adaptation process and target

comparison of training effect of various optimization

algorithms. In Figure 7, the ordinate is the loss Figure 8: Fitting effect of Nadam based salary forecasting model.
function (cost) and the abscissa is the period
(iteration). As can be seen from Figure 7, Adam and sample space is 100. Each algorithm is tested 10 times,
Nadam have better performance than other and the best test score is obtained. It can be observed
optimization algorithms in terms of convergence that the results of the Nadam optimized salary
speed, whereas Nadam is slightly better than Adam in prediction model in the test set is relatively higher as
terms of overall performance. compared to the other utilized algorithms. Finally, the
fitting effect of the Nadam salary prediction model is
To better reflect the performance of Nadam based
calculated and is shown in Figure 8.
salary prediction model, the comparison results
between Nadam and other ML regression algorithms It can be seen from Figure 8 that the prediction fitting
are given here. The size of the training sample space is effect of the Nadam based salary prediction model is
900, and the size of the test comparatively higher as compared to the rest of the
techniques. Although the normalization and
Comparision antinormalization of salary data in model training may
lead to some errors in calculation, this error is still
within the controllable range and has no great
influence on the final results of the model. Further, the
salary prediction model optimized by Nadam has a fast
convergence speed and good prediction effect and has

104
Scientific Programming 112

good accuracy as well, so it has certain practical and References

reference value. [1] Y. Lu, Research on the Application of Data Mining Technology
in Human Resource Management Systems, Beijing University
of Posts and telecommunications, Beijing, China, 2020.
4. Conclusion [2] L. Y. Wang and C. I. Peng, “Application of decision tree
algorithm in human resource recommendation technology,”
Globalization and worldwide economic integration Modern electronic technology, vol. 44, no. 3, pp.105–110,
have had a considerable impact on China’s economic 2021.
development, including the business community, in [3] T. Liu, M. Liu, H. Lihong, W. Luo, and W. Zheng, “Exploration
of human resource evaluation model based on ERP,” Energy
past few years. This results in disparities in many Technology, vol. 18, no. 9, pp. 90–93, 2020.
facets of business operations and trade, such as [4] Q. Xia and F. Huang, “Data-driven human resource
business thinking, observations, confidence, client management: review and prospect,” China personnel
science, vol. 12, no. 6, pp. 72–87, 2020.
psychology, behavioral culture, and salary prediction. [5] J. Tang, Programmer Recommendation System Based on
As a result, for sustainable and adaptive development GitHub, Yangzhou University, Yangzhou, China, 2020.
in the new context, each small- and medium-sized firm [6] Y. Li, S. Li, Y. song, F. Zhang, and X. Zhou, “Human Resource
Recommendation Based on K-means clustering algorithm of
must effectively leverage supportive solutions from spark platform,” Journal of Jinan University (Natural Science
the government, social media platforms, and small- & Medicine Edition), vol. 34, no. 5, pp. 430–435, 2020.
and medium-sized enterprises themselves. This study [7] L. Qi, Design and Implementation of a Personalized
Recommendation System for enterprise Training Resources,
mainly focuses on the application and implementation Harbin Institute of Technology, Harbin, China, 2018.
of the BP neural network-based salary forecasting [8] G. Liang, Human Resource Recommendation Algorithm Based
model. Multiple experiments have been performed by on Lifting Tree and Neural Network, South China University
of Technology, Guangzhou, China, 2018.
using a different number of neurons in the hidden
[9] C. Lin, Human Resource Recommendation Algorithm Based
layers and a different number of iterations in order to on Hybrid Genetic Ensemble Learning, South China University
check the stability and effectiveness of the proposed of Technology, Guangzhou, China, 2018.
[10] Z. Gu, Human Resource Recommendation Algorithm Based
model. This study also uses the gradient descent on the Hidden Semantic Model and Deep forest, South China
algorithm and its types by performing multiple University of Technology, Guangzhou, China, 2018.
experiments for each of its types. Among the [11] S.. Research, “On job recommendation of AI companies
based on data mining,” Value Engineering, vol. 36, no. 34, pp.
implemented types of gradient descent algorithm,
42–44, 2017.
Nadam showed good performance as compared to its [12] Z. Li, Human Resource Hybrid Recommendation Algorithm
other types. BP neural network-based salary Based on Spark, South China University of Technology,
forecasting model performed excellently by attaining Guangzhou, China, 2017.
[13] W. Chen, Research on Human Resource Recommendation
the training accuracy of 89.98% and validation Algorithm Based on Deep Learning, South China University of
accuracy of 84.05%. The proposed BP neural network- Technology, Guangzhou, China, 2017.
based salary forecasting model is anticipated to be of [14] F. Meng, Application Research of Human Resource
Recommendation Engine Based on Hybrid Recommendation
great help in human resource management systems. Algorithm, Beijing University of Technology, Beijing, China,
2014.
[15] P. N. Q. Huong, “The quality OF human resources IN small and
Data Availability medium enterprises: current situation, reason,
The data used to support the findings of this study are recommendations ON solutions,” Journal of Natural
Remedies, vol. 21, pp. 73–81, 2021.
included within the article. [16] D. L. Stone and J. H. Dulebohn, Emerging Issues in Theory and
Research on Electronic Human Resource Management
(eHRM), Elsevier, Amsterdam, Netherlands, 2013.
Conflicts of Interest [17] J. Heikkila, “A delphi study on e-HRM: future Directions,” in
The author declares that there are no conflicts of Proceedings of the 3rd European Academic Workshop on
Electronic Human Resource Management, pp. 229–249,
interest regarding the publication of this study.
Bamberg, German, May 2010.
[18] M. Berg, The Age of Manufactures, 1700-1820: Industry,
Innovation and Work in Britain, Routledge, Oxford, England,
2005.

105
Scientific Programming 113

[19] C. Trebilcock, The Industrialisation of the Continental Powers

1780-1914, Routledge, Oxford, England, 2014.
[20] F. W. Taylor, Scientific Management, Routledge, Oxford,
England, 2004.
[21] O. Yabanci, “From human resource management to
intelligent human resource management: a conceptual
perspective,” Human-Intelligent Systems Integration, vol. 1,
no. 2-4, pp. 101–109, 2019.
[22] A. Iqbal, “The strategic human resource management
approaches and organisational performance: the mediating
role of creative climate,” Journal of Advances in Management
Research, vol. 8, 2019.
[23] R. Mathis and J. Jackson, Human Resource Management
Mason, Thomson South-Western, Mason, OH, USA, 2008.
[24] M. Armstrong and S. Taylor, Armstrong’s Handbook of
Human Resource Management Practice, Kogan Page,
London, England, 2009.
[25] H. Francis and A. Keegan, “The changing face of HRM: in
search of balance,” Human Resource Management Journal,
vol. 16, no. 3, pp. 231–249, 2006.
[26] S. Strohmeier, “Research in e-HRM: review and implications,”
Human Resource Management Review, vol. 17, no. 1, pp. 19–
37, 2007.
[27] J. H. Marler and E. Parry, “Human resource management,
strategic involvement and e-HRM technology,” International
Journal of Human Resource Management, vol. 27, no. 19, pp.
2233–2253, 2016.
[28] L. Ma and M. Ye, “The role of electronic human resource
management in contemporary human resource
management,” Open Journal of Social Sciences, vol. 03, no.
04, pp. 71–78, 2015.
[29] S. Panos and V. Bellou, “Maximizing E-HRM outcomes: a
moderated mediation path,” Management Decision, vol. 54,
2016.
[30] M. Thite, e-HRM: Digital Approaches, Directions &
Applications, Routledge, Oxford, England, 2018.
[31] Y. Pan, “Heading toward artificial intelligence 2.0,”
Engineering, vol. 2, no. 4, pp. 409–413, 2016.
[32] H. Shah and K. Warwick, “Testing Turing’s five minutes,
parallel-paired imitation game,” Kybernetes, vol. 39, 2010.
[33] Y. Shiroishi, K. Uchiyama, and N. Suzuki, “Society 5.0: for
human security and well-being,” Computer, vol. 51, no. 7, pp.
91–95, 2018.
[34] S. I. Team, “Realization OF society 5.0 BY utilizing precision
agriculture into smart agriculture IN naro, Japan,” in
International Workshop on ICTs for Precision Agriculture, p.
58, Wuhan, China, 2019.
[35] P. Skobelev and S. Y. Borovik, “On the way from industry 4.0
to industry 5.0: from digital manufacturing to digital society,”
Industry, vol. 2pp. 307–311, Suceava, Romania, 2017.

106
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

PLOS ONE

Citation: Gong Y, Zhao M, Wang Q, Lv Z (2022) Design and interactive performance of human resource management system based on artificial
intelligence. PLoS ONE 17(1): e0262398. https:// doi.org/10.1371/journal.pone.0262398

Editor: Chi-Hua Chen, Fuzhou University, CHINA

Received: January 10, 2021

Accepted: December 23, 2021

Published: January 28, 2022

PeerReviewHistory: PLOS recognizes the benefits of transparency in the peer review process; therefore, we enable the publication of all of the content
of peer review and author responses alongside final, published articles. The editorial history of this article is available here:
https://doi.org/10.1371/journal.pone.0262398

Copyright:© 2022 Gong et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits
unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

DataAvailabilityStatement: All relevant data are within the manuscript and its Supporting Information files.

Funding: This research was supported in full by the National Natural Science Foundation of China (No. 61902203) and Key Research and
Development RESEARCH ARTICLE

Design and interactive performance of human resource management

system based on artificial intelligence
Yangda GongID1*, Min Zhao1, Qin Wang2, Zhihan Lv3
1 Business School, Hohai University, Nanjing, China, 2 Jiangsu Branch of China Mobile Group, Nanjing,
China, 3 School of Data Science and Software Engineering, Qingdao University, Qingdao, China

107
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

* [email protected]

Abstract
The purpose is to strengthen Human Resources Management (HRM) through information management using Artificial
Intelligence (AI) technology. First, the selection criteria of the applicant’s resume during recruitment and the formulation
standards of the contract salary are analyzed. Then, the resume information is extracted and converted into the data-type
format. Besides, the salary forecast model in the HRM system (HRMS) is designed based on the Back Propagation Neural
Network (BPNN), and network structure, parameter initialization, and activation function of the BPNN are selected and
optimized. The experimental results demonstrate that the algorithm optimized by the Nadm has shown improved convergence
speed and forecast effect, with 187 iterations. Moreover, compared with other regression algorithms, the designed algorithm
achieves the best test scores. The above results can provide references for designing the AI-based HRMS.

Introduction
With the socio-economic development, talents have become the mainstay of enterprise development, so more
enterprises are competing for human resources rather than the traditionally narrower sense of labor resources.
Furthermore, with structural informatization, enterprise data generation is increasing exponentially, which can no longer
be efficiently handled using the manual Human Resource Management System (HRMS) [1]. Accordingly, the Information
Management System (IMS) is employed to collect, store, manage, and analyze the exploding numbers of human
resource data to improve the traditional enterprise Human Resource Management (HRM) [2]. In other words, HRM and
Information Technology (IT) are combined in an attempt to establish an integrated HRMS to standardize the business
process of Human Resource (HR) departments, centralize HR information, and enhance HRM transparency [3]. The
quality of the HRMS will determine the enterprise performance and sustainable development. Therefore, it is urgent to
establish a high-performance HRMS.

Numerous studies have shown that HRM efficiency can be uplifted by integrating IT. Therefore, this paper utilizes big
data technology to manage and analyze HR data by designing a salary forecast model based on Artificial Intelligence (AI).
The proposed model forecasts the
Plan - Major Scientific and Technological Innovation Projects of ShanDong Province (2019JZZY020101).

Competinginterests: QW is a paid employee of the Jiangsu branch of China Mobile Group. This does not alter our adherence to PLOS ONE policies
on sharing data and materials. There are no patents, products in development or marketed products associated with this research to declare.

salary by analyzing job applicants’ resumes from basic information, school, society, and enterprise factors. Then, the
relevant parameters of the Back Propagation Neural Network (BPNN) are set and optimized to improve the forecast
accuracy of the proposed model, and the model performance is tested through experiment. This paper innovatively
integrates the BPNN into HRM to forecast employees’ salaries by analyzing their resumes to provide salary references
for the HR department for the applicants and to promote enterprise development.

108
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

Literature review
Abdullah et al. (2020) designed an enterprise cloud-based HRMS with 16 standard modules to solve HR problems using
the CodeIgniter Web framework, which was then launched and deployed on the Amazon Web Service elastic computing
cloud and used for an efficient enterprise HRM [4]. Necula and Strıˆmbei (2019) developed an architecture to
semantically enrich data through data science and semantic web technology for talent training. The experimental results
suggested that the classification effect of the proposed architecture was better than the commonly used regression
analysis, Random Forest (RF), and Support Vector Machine (SVM), and the proposed architecture could effectively mark
the resume data and use the semantic web to extract data information from the resume [5]. Jawad (2020) proposed a
website-based HRMS to manage employee activity information, such as salary, registration, and promotion. The HRMS
consisted of two parts: website design and database. Experimental results showed that the designed HRMS presented
high performance and efficiency in employee information storage and management [6]. Qin et al. (2020) put forward a
Recurrent Neural Network (RNN)-based applicant-job matching framework using job applicants’ perception ability,
word-level semantic representation, and experience. This method could reduce the dependence on manual labor and
improve employability. The information matching degree indicator was used to measure the importance of semantic
representation and the contribution of job experience to job requirements [7]. Serje et al. (2018) used the occupational
wage data from the International Labour Organization (ILO) to study econometric models of health workers’ income in
different countries. They employed the selection model to analyze the skill and income data of health workers. The
income of health workers varied in different countries and was negatively correlated with the country’s Gross National
Product (GNP). The results could predict the cost of health care intervention and resource needs during sustainable
development [8]. He et al. (2016) researched the cross-level relationship between salary differences and individual
turnover intention. They investigated and analyzed employees’ annual objective salaries and self-reporting attitudes
through the Questionnaire Survey (QS). Results demonstrated that the employee’s turnover intention and salary were
positively correlated: the lower the salary was, the stronger the turnover intention was [9]. Shen et al. (2019) explored
the organizational self-esteem, supervisory behaviors of communication between leaders, and restrictions on employee
feedback behavior. They analyzed the employment behavior of different managers and employees through the
hierarchical regression analysis and path analysis strategy. Results suggested that abuse supervision would directly
affect employees’ turnover intention [10].

Briefly, the current research on HRMS mostly focuses on employee recruitment, registration, and management.
Therefore, the salary forecast model is proposed for the BPNN-based HRMS to predict the salary of employees by
analyzing the resume of candidates, thereby providing a check and balance mechanism for applicants and the enterprise
interests.

Method
Salary forecast model in the BPNN-based HRM system
Usually, during the enterprise talent recruitment, qualified resumes are first picked
out according to applicants’ age, educational background, work experience, and
personal skills, which, traditionally, relies on manual selection. Thus, HR expertise
knowledge, as well as some industrial common senses, are often required from the
relevant personnel [11], where subjective human errors or prejudices might shut
some outstanding talents out of the enterprise threshold. Then, the applicants are
often informed of their possible salary range at the face-to-face interview stage,
which might vary significantly per department and position [12]. Yet, the final salary
is mostly determined by the department head after series of possible interviews

109
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

and discussions, which, again, involves tremendous manual works that might, in
turn, lead to deviations. There is an increasing voice calling for the elimination of
such non-objective factors under the current competitive market environment.

Meanwhile, under the traditional HRMS, given limited HR personnel and

countless applicants, resume selection is more a qualitative and speculative analysis
process than an objective and scientific evaluation procedure. AI technology can
well lend itself to such a predicament to help enterprises implement a salary
forecast model, which can predict applicants’ salary based on their resumes from as
early as the resume-selection stage and compares the forecast with their expected
salary to provide further references [13]. The model-forecasted salary can be used
as a benchmark salary for applicants, based upon which the actual salary can be
reasonably adjusted considering the specific departmental and positional standards.
An advantage of model-forecasted salaries is objectivity due to massive amounts of
data calculation; on the other hand, the salary forecast model substantially reduces
the workload of HR personnel and improves overall work efficiency [14].

The first step for salary forecast is the preprocessing of applicants’ resumes, after
which the extracted structured information is used to train the salary forecast
model against various data formats. The resume data can be classified as essential
information and supplementary information, as shown in Fig 1, in which the resume
is illustrated through hierarchically structured content. Common methods to extract
information reasonably include the rule-based extraction, the Cascaded Hybrid
Model, and the Conditional Random Field (CRF) extraction method.

Employee salary determination is an intricate business that involves employee

job specialty, as well as some subjective and objective factors, such as the nature of
the enterprise. So is the salary forecast process that depends on the completeness
of the resume information and the model performance to extract data features
accurately and efficiently. The resume features division based on its content and
formats read:
(1) Personal factors: name, age, gender, phone number, and other information,
divided according to the degree of association with salary forecast. For example,
residence and birthplace will affect employment tenure; name, phone number,
and email address are unique identifiers of the applicants; there may be special
requirements on age and gender for specific positions [15].

(2) School-related factors including school-time, alma mater, educational

background, specialty, and awards. The alma mater and specialty can reflect an
applicant’s learning ability and expertise skills, which also affect the salary
forecast; awards are the manifestation of the applicant’s school performance
and learning attitudes.

110
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

(3) Social factors including served companies, positions, work hours, and length of
service. Served companies and positions can reflect the level of personal
abilities and skills. Length

level

https://doi.org/10.1371/journal.pone.0262398.g001

of service can reflect the mastery of relevant skills and personal adaptability.
The number of served companies also affects the result of the salary forecast
[16].
(4) Enterprise factors are more closely related to the target enterprises. They are
not included in the resumes, including the applying positions and occupation
levels. Different enterprises in the same industry have different salary plans. The
position applied to is the core of the salary forecast, which has a significant
impact on the salary level. The occupational level is the requirement of salary
level and the key to salary forecast [17].

111
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

Therefore, the salary demands of applicants should be forecasted from multiple

aspects. Afterward, the character data extracted from the resume are converted
into numerical data for subsequent data preprocessing and model training.

The salary forecast is a regression analysis process. The Neural Network (NN) in
Machine Learning (ML) can be used as a salary regression forecast model, which is a
parallel interconnected network composed of adaptive neural units and simulates
the interaction process between the biological nervous system with the outside
world. BPNN is the most common NN model. Theoretically, a 3-layer BPNN can
approach a continuous function of arbitrary precision with a definite learning ability
[18]. Yet, the network structure, parameter settings, and optimization algorithms
will affect the NN training results and the salary model’s forecast effect. The
learning process of the NN is the adjustment process of neuron parameters. BPNN’s
error backpropagation process can adjust parameters. BPNN adopts a Gradient
Descent (GD) strategy, which uses the target’s negative gradient direction as the
search direction. During continuous iterations, the network parameters are
updated, the model converges, and finally, the network parameters are adjusted
[19].

BPNN contains a multi-layer signal feedforward NN, and the learning process is
divided into two processes: forward propagation and directional propagation. The
sample data enters the network from the input layer during the forward
propagation, processed by the hidden layer and output from the output layer [20].
The error backward propagation process starts from the output layer, returns the
results by layers in the reverse direction and the connection of neurons, and adjusts
the neuron parameters in the path [21]. The processes of BPNN’s forward and
backward propagation are illustrated in Fig 2.

In the signal forward propagation process of the 3-layer BPNN, xi represents the
output of the input layer neuron, lj denotes the output of the hidden layer neuron,
and zk indicates the output of the output layer neuron; αj refers to the threshold of
the hidden layer, and βk stands for the threshold of the output layer; wij indicates
the weight of the input layer to the hidden layer, and vjk denotes the weight of the
hidden layer to the output layer. The sample data are not processed in the input
layer, and the output lj of the j-th node in the hidden layer can be expressed as Eq
(1).

X
lj ¼ fðnetjÞ ¼ fð ðwijxi ajÞÞ ð1Þ
i

The output zk of the k-th node in the output layer reads:

X zk ¼
fðnetkÞ ¼ fð ðvjklj bkÞÞ ð2Þ
j

112
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

The output zk of the output layer is the calculation of forward propagation.

However, a gap exists between the generated forecast value and the actual value.
The loss function E can serve

https://doi.org/10.1371/journal.pone.0262398.g002

as a standard to measure the error. The expression of loss function E reads:

E ¼ Cðy;zÞ ð3Þ

In Eq (3), θ = θ(α, β, w, v) represents all the network parameters, and z =

(z1,z2,...zk) denotes the output vector [22]. BPNN uses a GD method to minimize the
loss function in the continuous iteration process [23]. The calculation process is
described as follows.

First, the connection weight gradient @@vEjk of the output layer is calculated
according to:

113
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

@E ¼ @E @netk ¼ ð4Þ
@E @zk @netk
@vjk @netk @vjk @zk @netk @vjk

@E 0 0

¼ C ðy;zÞjz f ðnetkÞlj ð5Þ

@vjk k

The output error δk, weight gradient @@vEjk, and threshold gradient @@bEk of the

output node are calculated as in Eqs (6)–(8), respectively:

dk ¼ C0ðy;zÞjzk f 0ðnetkÞ ð6Þ

@E
¼ dk lj ð7Þ
@vjk

@E

¼ dk ð8Þ
@bk

The connection weight gradient of the hidden layer @@wEij can be presented as:

@E @E @netj @E @lj @netj @wij

@netj @wij @lj @netj @wij

¼
¼ ð9Þ
@E X 0 0 0

@ wij ¼ k C ðy;zÞjzk f ðnetkÞvjkf ðnetjÞxi ð10Þ

@E X 0 0 X
¼ d v f ðnet Þx ¼ f ðnet Þx dv ð11Þ
@wij k k jk j i j i k k jk

114
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

The calculations of the output error γj, weight gradient @@wEij, and threshold

gradient @@aEj of the hidden layer read:

X
gj ¼ f ðnetjÞ
0 dkvjk ð12Þ
k

@E
¼ gjxi ð13Þ
@wij

@E

¼ gj ð14Þ
@aj

Meanwhile, the learning rate η of each network layer can be adjusted to

optimize the update step size and convergence speed of the iterative algorithm
[24]. In BPNN, parameters are adjusted according to the negative gradient direction
of the target. Assuming that the learning rate of each network layer is the same,
then:
@E

Dwij ¼ Z ¼ Zgjxi ð15Þ

@wij

@E

Daj ¼ Z ¼ Zgj ð16Þ

@yj

@E

Dvjk ¼ Z ¼ Zdklj ð17Þ

@vjk

@E

Dbk ¼ Z ¼ Zbk ð18Þ @bk

X
gj ¼ f ðnetjÞ
0 dkvjk ð19Þ
C0ðy;zÞjzk f 0ðnetkÞ ð20Þ

115
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

In Eqs (19) and (20), f() represents the activation function, and c() refers to the
error function. Therefore, the gradient of the hidden layer parameters and the
output layer parameters can be calculated by layers. This process is called
backpropagation [25].

Determining parameters of the salary forecast model

(1) Network structure. Before BPNN can approach non-linear functions with
arbitrary precision through its non-linear mapping, generalization, and fault
tolerance capabilities, the network structure, activation function, and parameter
initialization must be solved [26]. Since the salary forecast model processes
multiple inputs while producing only one output, the number of input neurons
must be determined. Based on the experimental analysis, the number of input
and output neurons is set to 14 and 1, respectively. Since the hidden layer does
not have a definite calculation equation, the hidden layer number is estimated
first then determined through trial and error. The number of hidden layer
neurons nh can be determined according to:
pffiffiffiffiffiffiffiffiffiffi
ð21Þ
nh ¼ nino
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
nh ¼ ni þ no þ k;k 2 ½1;10 ð22Þ
In Eqs (21) and (22), ni represents the number of nodes in the input layer, and no
denotes the number of nodes in the output layer [27].
(2) Activation function. The activation function contains linear or non-linear
functions used to solve linear problems and uncertain problems, respectively.
The commonly used activation function Sigmoid function is expressed as Eq
(23).
1
fðxÞ ¼ 1þex ð23Þ

The output result of this function is between (0, 1), and the function has
differentiability and saturated nonlinearity, which can enhance the non-linear
mapping ability of the network [28].
(3) Parameter initialization. Parameter initialization will affect the training results
and the convergence degree of the model. If the parameter setting is
unreasonable, the model will fall near the local minimum during training and
fails to converge. In the initial situation, the initial connection weight is
accumulated, and the state of each neuron is 0. Neurons in the same layer

116
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

(4) should not be assigned the same weight; otherwise, the calculation result will
be the same output [29]. Therefore, it is necessary to ensure that the
connection weight is a random decimal number, and the value difference
between each other is small. Hence, a good convergence speed will be
obtained. This paper selects a Gaussian random number
pffiffiffiffiffi
with a mean value of 0, and a standard deviation of nin (nin represents the
number of neuron input connection weights), as well as a Gaussian random
number with a mean value of 0, and a standard deviation of 1 as the weight and
threshold parameters, respectively [30]. During algorithm training, the weight
of the corresponding NN is reduced to minimize the recruitment discrimination
caused by gender, age, and other factors.
(5) Loss function. It can measure the error between the NN forecasted value and
the actual value, which is the important criterion for training the learning
model. Here, the objective function is calculated by quadratic Mean Square
Error (MSE), as shown in Eq (24).

E¼ ðy zÞ2 ¼ Cðy;zÞ ð24Þ

In Eq (24), y represents the objective function, and z denotes the output vector
[31].
(6) Optimization methods. BPNN has a simple structure and strong learning ability,
but its complexity and performance are affected by the parameter initialization
and network structure. Besides, given a flat error surface, oscillation will occur
and affect the convergence speed. During the GD, the model easily falls into a
local minimum and affects its forecast result. In the actual application, the
convergence speed of BPNN needs to be optimized to avoid falling into a local
minimum [32]. Common optimization methods include:
1) In BPNN, the GD algorithm is applied to update the parameters in the
backward propagation. According to the data bulk, the GD algorithm can be
divided into Batch Gradient Descent (BGD) [33], Stochastic Gradient Descent
(SGD) [34], and Mini-batch Gradient Descent (MBGD) [35].

2) The last parameter change and the current parameter increment are
connected by adding the former to the latter. This method of adding
momentum to the parameter increment is called the additional momentum
method, including the Momentum method and Nesterov Accelerated
Gradient (NAG) method [36].

117
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

3) The adaptive learning rate method dynamically adjusts the learning rate
during the model training process to speed up convergence and reduce
oscillations. Common adaptive learning rate optimization methods are
Adagrad and RMSprop [37].

4) Combining the additional momentum method and the adaptive learning rate
method can obtain a hybrid optimization method that can optimize the GD.
Such hybrid method has a fast convergence speed and fewer oscillations,
including Adaptive Moment Estimation (Adam) and Nesterov-accelerated
Adaptive Moment Estimation (Nadam) [38].

In summary, the ratio of the number of neurons in the input layer, the hidden layer,
and the output layer is 14:15:1 in the designed 3-layer BPNN. The activation
function is a Sigmoid function, while the error function is a quadratic MSE function.
MBGD and Nadam optimization algorithms are used to optimize the model, and the
BPNN parameters and experimental environment are displayed in Tables 1 and 2.
Therefore, the output errors δk and γj of the output layer and the hidden layer are
calculated according to:

dk ¼ ðyk zkÞzkð1 zkÞ ð25Þ

X
gj ¼ ljð1 ljÞ dkvjk ð26Þ
k

In Eqs (25) and (26), yk and zk represent the target value and output value of the
k-th node.

The salary forecast model’s training process based on BPNN is shown in Fig 3 [39].

Introduction to experimental samples

Because salary forecast belongs to a multiple-input-single-output mapping process,
effective salary forecast results can be obtained by considering different influencing
factors of salary like the model’s output and optimizing the influence weight of each
factor in the model through training. The information in the resume library of an
enterprise is analyzed to verify the effect

network
number neurons
by neurons
neurons

Momentum

118
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

https://doi.org/10.1371/journal.pone.0262398.t001

of the designed model. The number of samples is 1,000, of which the number of
recruits is 298, the number of rejections is 702, and each resume has a
corresponding salary status. Data characteristics of the resume include age, salary,
gender, the highest academic credential, major, marital status, job position,
position applied, work experience, and length of service. Since the salary forecast
model structure is affected by the training samples and model parameters, the
most suitable parameter settings are obtained by continuously adjusting the model
parameters. Recruitment bias caused by factors such as gender and age may arise
during employee recruitment. However, in the present work, the resume data of
the enterprise after successful recruitment is used so that the work does not
directly consider the possible moral hazard. Among the risks, the generated salary
forecast model also meets the enterprise’s recruitment requirements. Therefore,
the research results do not consider the possible moral hazard. To better compare
the classification effect of the classifier, the following indicators are employed to
evaluate the classification effect of the classifier. Among them, TP refers to the
proportion of resumes correctly classified as entry, FP denotes the proportion of
resumes incorrectly classified as entry, TN indicates the proportion of resumes
correctly classified as non-entry, and FN represents the proportion of resumes
incorrectly classified as the entry.
(1) Precision refers to the total proportion of positive samples correctly classified by
the classifier.
TP TP þ FP

Precision ¼ ð27Þ
(2) Recall in Eq (28) indicates the proportion of positive samples that are correctly
predicted.
TP TP þ FN

Recall ¼ ð28Þ

Results and discussion

Parameter selection
According to the empirical equation, the number of neurons is in the range of [5,
14]. Hence, the model with several neurons between [3, 16] is trained and tested.
The number of samples used is 1,000, with 500 in the training set and 500 in the
validation set. The hidden layer neurons are trained 30 times to obtain the average
result. The number of iterations is set at 1,000, with the training loss accuracy
E<0.01. The training results are displayed in Fig 4.

119
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

According to Fig 4, when the number of neurons in the hidden layer is 14, the
model’s training loss is the smallest, which is 0.009914. When the number of
neurons is 15, the number of iterations and the model’s verification loss are the
smallest, of 64 and 0.010511, respectively. Therefore, the number of nodes in the
input layer, the hidden layer, and the output layer of the designed BPNN is 14, 15,
and 1, respectively. Such a structure can accelerate the model’s running speed and
reduce the verification loss and training loss.
Table 2. Experimental hardware and software environment.
environment @2.40GHz environment

memory
https://doi.org/10.1371/journal.pone.0262398.t002

process forecast
https://doi.org/10.1371/journal.pone.0262398.g003

The effect of the optimization algorithm on the model is verified through the
Leave-OneOut (LOO) method. The sample-set with 1,000 samples is divided into a
training set with 900 samples and a test set with 100 samples. The number of cycles
is set to 10,000, or the training is stopped when the training loss accuracy E is less
than 0.005. The convergence speed of the model optimized by different algorithms
is shown in Fig 5; the comparison of training results is displayed in Fig 6.

As shown in Fig 5, among various optimization algorithms, Adam and Nadm have
faster convergence speeds than other optimization algorithms, which can quickly
converge to the minimum and reduce the model’s running time. Furthermore, the

120
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

convergence effect of Nadm is better than that of Adam. According to the

training results comparison of six algorithms, the test scores and training losses of
various optimization algorithms are close. However, there are noticeable
differences in the number of training cycles. SGD has the largest

Fig 4. Comparison of training results of different numbers of neurons in the hidden layer.

https://doi.org/10.1371/journal.pone.0262398.g004

number of training cycles, of 3,228 times. In contrast, Nadm has the minimal
training cycles of 187 times. Therefore, among the various NN GD algorithms, the
hybrid optimization algorithm Nadm presents the best optimization effect and
convergence speed. Hence, Nadm is chosen as the GD optimization algorithm for
the salary forecast model.

Analysis of salary relevance

It has been argued that salary level might somehow reflect the relevance between
talent supply and recruitment positions. This section selects 10,000 pieces of
enterprise recruitment data, including the lowest salary, highest salary, and average
salary, location, company size,

121
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

methods.

https://doi.org/10.1371/journal.pone.0262398.g005

Fig 6. Comparison of training results of different optimization methods.

https://doi.org/10.1371/journal.pone.0262398.g006

financing, education level, work experience, and job type. First, the relevance
among different influencing factors is analyzed using the Pearson correlation
coefficient method, and the results are shown in Fig 7. Fig 7 reveals that the
relevance between salary is the strongest, and the relevance among education
level, job type, and work experience and salary is also stronger than other
influencing factors. Therefore, salary is taken as the dependent variable with the

122
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

influencing factors including job type, work experience, and education level as the
independent variable, and the analysis results are shown in Fig 8.

https://doi.org/10.1371/journal.pone.0262398.g007

Fig 8. The relationship between dependent variables and salary. a) Job type; b) Work experience; c) Education level.

123
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

https://doi.org/10.1371/journal.pone.0262398.g008

Fig 8A suggests that there is a certain relationship between job type and salary.
The salary of different jobs varies much. The salary range of data risk jobs and
products is 15-25k/month. The development of the Internet in recent years has
made the income of related industries continue to rise. Operation positions and
design positions are affected by personnel, and the salary will also change. Fig 8B
indicates that the salary will continue to improve with the accumulation of work
experience. The salary will not change much within three years, but will increase
after three to five years, and will increase significantly after five to ten years. The
maximum monthly salary can reach more than 40 k/month. Hence, after several
years of work experience, employees can get a higher salary. Fig 8C shows that with
the improvement of education level, the salary also rises. The salary of
undergraduates is 50% higher than that of college students. The salary of masters
and doctors is much higher than that of other education levels, and the salary
ceiling of doctors is higher. To sum up, job type, work experience, and education
level are the main factors that affect salary, and the different influencing factors will
also affect salary. But in general, the higher the education level is, the richer the
work experience is, and the more likely the employees in popular positions are to
get a high salary.

Performance test
The resume information status of the experimental sample is summarized in Fig 9.
The salary forecast model based on BPNN is applied to fit the salary data to verify
its actual forecast effect, and the results are shown in Fig 10. Common regression
algorithms are adopted for comparative simulations, including Linear, Polynomial,
Ridge, Lasso, and ElasticNet. The ratio of the correct number of samples output by
different models to all samples is taken as the model’s score. The scores of the
training process and testing process are presented in Fig 11. The

a)
https://doi.org/10.1371/journal.pone.0262398.g009

124
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

forecast

https://doi.org/10.1371/journal.pone.0262398.g010

number of samples in the training set used for model training is 900, and the
number of samples in the test set is 100. Six experiments are carried out on each
model and the average value is taken as the final result.

Fig 9 presents that the recruitment data involves different job types, and the data
content includes salary level, gender, education level, and other information.
According to Fig 10, the BPNN-based salary forecast model optimized by the Nadm
gradient has an excellent fitting performance. However, some errors still exist. The
reason may be that the data are normalized during the model training process,
resulting in errors in the calculation results. However, the overall error is
acceptable. As shown in Fig 11, the Nadm gradient optimization model has the
highest test score than other algorithms.

forecast

125
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

https://doi.org/10.1371/journal.pone.0262398.g011

Performance comparison
BPNN is employed to analyze the establishment information to obtain the
forecasted salary and compare the salary level of the sample database from
precision and recall rate. The forecast results of the LSTM network, GRU network,
Bi-LSTM network, and Bi-GRU network are compared, as illustrated in Fig 12.

As shown in Fig 12, the precision and recall rate of BPNN optimized by the GD
algorithm for salary forecast is above 0.9, which is better than the forecast results
of the LSTM network, GRU network, Bi-LSTM network, and Bi-GRU network. Hence,
the actual effect of the designed salary forecast model can provide better accuracy
than similar models. Thus, during the actual application, a desirable salary can be
provided to applicants by analyzing their resumes. The results can provide a
theoretical reference for establishing a reasonable employee salary system.

In summary, the designed salary forecast algorithm based on the BPNN model
uses a network structure with 14:15:1 (input layer: hidden layer: output layer) of
the number of neurons. Besides, the Nadm gradient optimization algorithm is used
to optimize the model to get a faster convergence speed and excellent forecast
effect. Compared with other regression algorithms, the algorithm used here has the
best forecast effect and test score. The actual forecast results also show that the
salary forecast model can provide desirable salaries for applicants. Therefore, the
designed salary forecast algorithm can apply to forecast the salary in the HRM
system.

Conclusion
This work aims to build an intelligent HRMS and improve the efficiency of HRM. A
HumanComputer Interaction (HCI)-based HRMS is designed using AI technology to
strengthen enterprises’ management and development capabilities. First, J2EE is
employed to design a modularized HRMS. Second, the Artificial Neural Network
(ANN) is adopted to optimize the HRMS, and the salary forecast module is
implemented to effectively judge the applicants’ ability according to the resume
and industry information, thereby offering a reasonable salary.

Eventually, the performance of the designed salary prediction model of the HRMS is
tested

126
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

forecast

https://doi.org/10.1371/journal.pone.0262398.g012

and analyzed. The experimental results demonstrate that the network structure,
parameter settings, and gradient optimization algorithm will affect the model’s
forecast structure. Tests prove that the Nadm gradient optimization algorithm can
effectively improve the model’s convergence speed and actual fitting effect.
Compared with other algorithms, the model optimized by Nadm has optimal test
scores. Therefore, the proposed algorithm can be applied to the salary forecast of
the HRM system. However, there are still some shortcomings. The salary of
applicants might be affected by many factors in real life, but this paper conducts the
correlation analysis only on several main influencing factors, so the proposed salary
forecast model is relatively simple, with a weak data feature processing ability. In
the research process, the model has been optimized to further reduce the impact of
potential recruitment discrimination on the output of the model. In the follow-up
research, it is worth considering more possible influencing factors to enhance the
model forecast ability and accuracy.

Supporting information
S1 Data.
(ZIP)

Author Contributions
Conceptualization: Yangda Gong, Min Zhao, Zhihan Lv.
Data curation: Yangda Gong, Min Zhao, Zhihan Lv.
Formal analysis: Yangda Gong, Min Zhao, Zhihan Lv.
Funding acquisition: Yangda Gong, Zhihan Lv.

127
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

Methodology: Min Zhao, Zhihan Lv.

Project administration: Min Zhao.
Resources: Yangda Gong, Min Zhao, Zhihan Lv.
Software: Yangda Gong, Min Zhao.
Supervision: Min Zhao.
Visualization: Zhihan Lv.
Writing – original draft: Yangda Gong, Qin Wang.
Writing – review & editing: Yangda Gong, Min Zhao, Qin Wang, Zhihan Lv.

References
1. Sodhro AH, Pirbhulal S, Luo Z, et al. Towards optimal resource management for IoT based Green
and sustainable smart cities. J Clean Prod, 2019; 220:1167–1179.
2. Hong CH, Varghese B. Resource management in fog/edge computing: a survey on architectures,
infrastructure, and algorithms. Acm Comput Surv, 2019; 52(5): 1–37.
3. Raut RD, Luthra S, Narkhede BE, et al. Examining the performance-oriented indicators for
implementing green management practices in the Indian agro sector. J Clean Prod, 2019; 215:926–
943.
4. Abdullah PY, Zeebaree SRM, Jacksi K, et al. An hrm system for small and medium enterprises (SME)
s based on cloud computing technology. Int J Found Comput S, 2020; 8(8):56–64.
5. Necula SC, Strıˆmbei C. People Analytics of Semantic Web Human Resource Re´sume´s for
Sustainable Talent Acquisition. Sustainability-Basel, 2019; 11(13):3520.
6. Jawad WK. Design and Implementation of E-Human Resource Management System for IT Company.
Int J Found Comput S, 2020; 3(1):1–6.
7. Qin C, Zhu H, Xu T, et al. An enhanced neural network approach to person-job fit in talent recruitment.
ACM T Inform Syst Se, 2020; 38(2):1–33.
8. Serje J, Bertram MY, Brindley C, et al. Global health worker salary estimates: an econometric analysis
of global earnings data. Cost Effect Resour A, 2018; 16(1):1–9. https://doi.org/10.1186/s12962-
0180093-z PMID: 29541000
9. He W, Long LR, Kuvaas B. Workgroup salary dispersion and turnover intention in China: A contingent
examination of individual differences and the dual deprivation path explanation. Hum Resour
ManageUs, 2016; 55(2):301–320.
10. Shen C, Yang J, He P, et al. How does abusive supervision restrict employees’ feedback-seeking
behavior? J Manage Psychol, 2019; 34(8):546–559.
11. Boon C, Den Hartog D N, Lepak D P. A systematic review of human resource management systems
and their measurement. J Pain Symptom Manag, 2019; 45(6):2498–2537.
12. Yong JY, Yusliza MY, Ramayah T, et al. Nexus between green intellectual capital and green human
resource management. J Clean Prod, 2019; 215:364–374.
13. Singh SK, Del Giudice M, Chierici R, et al. Green innovation and environmental performance: The
role of green transformational leadership and green human resource management. Technol Forecast
Soc, 2020; 150:119762.
14. Macke J, Genari D. Systematic literature review on sustainable human resource management. J
Clean Prod, 2019; 208:806–815.
15. Chams N, Garcı´a BJ. On the importance of sustainable human resource management for the
adoption of sustainable development goals. Resour Conserv Recy, 2019; 141:109–122.

128
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

16. Roscoe S, Subramanian N, Jabbour CJC, et al. Green human resource management and the
enablers of green organisational culture: Enhancing a firm’s environmental performance for
sustainable development. Bus Strateg Environ, 2019; 28(5):737–749.
17. Pham NT, Tučkova´ Z, Jabbour CJC. Greening the hospitality industry: How do green human resource
management practices influence organizational citizenship behavior in hotels? A mixed-methods
study. Tourism Manage, 2019; 72:386–399.
18. Saeed BB, Afsar B, Hafeez S, et al. Promoting employee’s proenvironmental behavior through green
human resource management practices. Corp Soc Resp Env Ma, 2019; 26(2):424–438.
19. Kim YJ, Kim WG, Choi HM, et al. The effect of green human resource management on hotel
employees’ eco-friendly behavior and environmental performance. Int J Hosp Manag, 2019; 76:83–
93.
20. Lee HW, Pak J, Kim S, et al. Effects of human resource management systems on employee proactivity
and group innovation. J Pain Symptom Manag, 2019; 45(2):819–846.
21. Ferraris A, Erhardt N, Bresciani S. Ambidextrous work in smart city project alliances: unpacking the
role of human resource management systems. Int J Hum Resour Man, 2019; 30(4):680–701.
22. Ogbonnaya C, Messersmith J. Employee performance, well-being, and differential effects of human
resource management subdimensions: Mutual gains or conflicting outcomes? Ger J Hum Resour
Man, 2019; 29(3):509–526.
23. Peccei R, Van De Voorde K. The application of the multilevel paradigm in human resource
management–outcomes research: Taking stock and going forward. J Pain Symptom Manag, 2019;
45(2):786– 818.
24. Jabbour CJC, Sarkis J, de Sousa Jabbour ABL, et al. Who is in charge? A review and a research
agenda on the ‘human side’of the circular economy. J Clean Prod, 2019; 222:793–801.
25. Barrena-Martı´nez J, Lo´pez-Ferna´ndez M, Romero-Ferna´ndez PM. Towards a configuration of
socially responsible human resource management policies and practices: Findings from an academic
consensus. Int J Hum Resour Man, 2019; 30(17):2544–2580. Doi: 10.1080/09585192.2017.1332669
26. Parry E, Battista V. The impact of emerging technologies on work: a review of the evidence and
implications for the human resource function. Forsch Komplementmed, 2019; 1(5):5.
27. Chou YC, Yen HY, Dang VT, et al. Assessing the human resource in science and technology for
Asian countries: Application of fuzzy AHP and fuzzy TOPSIS. Symmetry-Basel, 2019; 11(2):251.
28. Drobyazko S, Barwińska-Małajowicz A, Ślusarczyk B, et al. Innovative entrepreneurship models in
the management system of enterprise competitiveness. Entrep Res J, 2019; 22(4):1–6.
29. Wang T, Liang Y, Jia W, et al. Coupling resource management based on fog computing in smart city
systems. J Netw Comput Appl, 2019; 135:11–19.
30. Namany S, Al-Ansari T, Govindan R. Sustainable energy, water and food nexus systems: A focused
review of decision-making tools for efficient resource management and governance. J Clean Prod,
2019; 225:610–626.
31. Siriyanun S, Mukem A, Jermsittiparsert K. Bridging Workplace Democracy and Supply Chain
Integration through High Involvement of Human Resource Practices. Supply Chain Manag, 2019;
8(5):685– 696.
32. Stachova´ K, Papula J, Stacho Z, et al. External partnerships in employee education and development
as the key to facing industry 4.0 challenges. Sustainability-Basel, 2019; 11(2):345.
33. Purwanto A, Asbari M, Budi Santoso P. Influence of Transformational and Transactional Leadership
Style toward Food Safety Management System ISO 22000: 2018 Performance of Food Industry in
Pati Central Java. Sustainability-Basel, 2019; 7(2):180–185.
34. Tambe P, Cappelli P, Yakubovich V. Artificial intelligence in human resources management:
Challenges and a path forward. Calif Manage Rev, 2019; 61(4):15–42.
35. Cooke FL, Cooper B, Bartram T, et al. Mapping the relationships between high-performance work
systems, employee resilience and engagement: A study of the banking industry in China. Int J Hum
Resour Man, 2019; 30(8):1239–1260.
36. Santana M, Morales-Sa´nchez R, Pasamar S. Mapping the Link between Corporate Social
Responsibility (CSR) and Human Resource Management (HRM): How Is This Relationship
Measured? Sustainability-Basel, 2020; 12(4):1678.

129
PLOS ONE Design and interactive performance of human resource management system based on artificial intelligence

37. Gyura´k Babelˇova´ Z, Stareček A, Koltnerova´ K, et al. Perceived Organizational Performance in

Recruiting and Retaining Employees with Respect to Different Generational Groups of Employees
and Sustainable Human Resource Management. Sustainability-Basel, 2020; 12(2):574.
38. Chae C, Yim JH, Lee J, et al. The Bibliometric Keywords Network Analysis of Human Resource
Management Research Trends: The Case of Human Resource Management Journals in South
Korea. Sustainability-Basel, 2020; 12(14):5700.
39. Bagherinezhad Z, Abdullahi B, Hassanpour A, et al. Designing a Model for Social Responsibility
Human Resource Management in Electric Power Industry. J Educ Comput Res, 2020; 12(2):147–
187.

130