2020 International Conference on Advance in Ambient Computing and Intelligence (ICAACI)

Model and Implementation of E-commerce

2020 International Conference on Advance in Ambient Computing and Intelligence (ICAACI) | 978-1-7281-9753-1/20/$31.00 ©2020 IEEE | DOI: 10.1109/ICAACI50733.2020.00050

Recommendation System Based on User Clustering

Jiaoli Zhou, Fang Wan*, Ru Jing

Hainan Vocational University of Science and Technology
Haikou, Hainan 571126, China

Abstract—The growth of data volume provides more product recommendation service for each customer [5].
opportunities for the development of e-commerce, but it also
produces the problem of information overload. That is, users With the increasing scale of e-commerce, the categories
can't quickly and accurately obtain the really interesting and quantities of commodities are becoming more and more
demand information from massive commodity information. complicated. The recommendation system can simulate the
The existing e-commerce website search systems and search salespeople in stores, and according to the interests and
engines mainly filter information through the search conditions hobbies of users, recommend products that meet their
input by users, without personalized consideration, and cannot interests and hobbies to the target users [6]. It is a new
provide convenient, efficient and accurate services according to opportunity and a new challenge for the merchants that the
the needs of users. Collaborative filtering is the most widely two sides of e-commerce transaction do not meet each other
used and successful recommendation technology in and complete each link of the traditional business model
recommendation system. However, with the increase of the through the Internet. The application of recommendation
number of users and commodities in e-commerce system, the system can not only provide personalized service for users,
time consumption of searching the nearest neighbor of the but also establish a long-term and stable relationship with
target user in the whole user space also increases dramatically, users, improve user loyalty, prevent users from losing, and
which leads to the performance degradation of the system. In finally increase the sales of e-commerce websites [7]. Under
this paper, a collaborative recommendation method based on the increasingly fierce competition environment, the e-
user clustering is proposed. Firstly, the user's score on the
commerce recommendation system can effectively retain
commodity category is calculated, and then the user is
clustered based on the user's score on the commodity category.
users, prevent users from losing, change the viewers of e-
commerce websites into buyers, improve the cross-selling
Keywords—E-commerce, User clustering, Recommendation ability of e-commerce websites, and improve customers'
system loyalty to e-commerce websites, thus improving the sales of
e-commerce systems [8]. With the rapid increase of the
I. INTRODUCTION number of users and commodities, the time required to
Under the trend of global economic integration, search the nearest neighbor in the whole user space also
informatization and networking, China's computer increases rapidly, and it is increasingly difficult to meet the
ownership, the number of Internet users, and the number of requirements of real-time response of users [9]. In this paper,
websites have grown rapidly, providing favorable conditions a collaborative recommendation implementation method
for the widespread development of e-commerce, and e- based on user clustering is proposed. Firstly, the user's
commerce has created huge business opportunities [1]. With rating of commodity category is calculated, and then the
the popularity of the Internet and the development of e- user is clustered based on the user's rating of commodity
commerce, while the e-commerce system provides users category.
with more and more choices, its structure has become more II. ARCHITECTURE OF PERSONALIZED RECOMMENDATION
complicated, and users are often lost in a large number of
SYSTEM FOR E-COMMERCE
commodity information spaces and cannot find themselves
smoothly Required goods [2]. Faced with a large amount of With the arrival of the e-commerce era, the domestic
product information, it is difficult for consumers to quickly personalized recommendation system is in a stage of
and accurately select the products that really suit their needs. blowout development. The origin of the recommendation
The e-commerce recommendation system is to provide users system can be traced back to extensive work in other related
with product information and suggestions based on fields, such as cognitive science, approximation theory,
understanding and learning customer needs and preferences, information retrieval, prediction theory, and even with
and recommend products that may be of interest to users [3]. management science. With the popularity of the Internet and
Providing personalized services has become one of the the development of social networks, the recommendation
important issues that need to be resolved to further improve system has gradually become an important research content
the quality of online content services, and is also a in social networks. Various types of massive data are stored
development direction for future network content services in the network, and it becomes more and more difficult to
[4]. Recommendation system has a good development and obtain useful information effectively. Although there are
application prospect in e-commerce system, but with the many search engines on portals, the degree of
further expansion of e-commerce system, e-commerce personalization is limited, and they often cannot meet
recommendation system also faces the problem of how to specific user needs. The recommendation system
improve accuracy and timeliness to provide personalized recommends products and information to the user based on

978-1-7281-9753-1/20/$31.00 ©2020 IEEE 197

DOI 10.1109/ICAACI50733.2020.00050

Authorized licensed use limited to: ULAKBIM UASL - YILDIZ TEKNIK UNIVERSITESI. Downloaded on May 14,2021 at 18:09:57 UTC from IEEE Xplore. Restrictions apply.
 the user's purchase history and interest characteristics. recommendation system is to recommend items similar to
Personalized recommendations are based on data mining those he liked in the past. User clustering is a label model
and information acquisition. When a user visits an e- abstracted by comprehensively analyzing user data, which
commerce website, the recommendation system provides can represent all dimensions of users, among which
services for the user's shopping and accurately recommends dimensions generally include demographic dimension,
suitable products for him [10]. If the user data has fewer interest dimension and business dimension. The basic idea
content features or less useful features, the content-based of content-based recommendation system is to count the
recommendation system cannot play a better items that the user once liked or bought according to the
recommendation effect. Usually the model is already existing user data, and then recommend similar items to the
established and is obtained from the training data. Or define user.
a classification function and divide the data into pre-defined
types through the classification function that has been III. PERSONALIZED EECOMMENDATION PPROCESS OF E-
trained. COMMERCE RECOMMENDATION SYSTEM

The method of measuring the similarity between users In the age of booming social media, the lack of data for
and is as follows. First, all items rated by users i and j are credit assessments has begun to change. The ever-growing
obtained, and then the similarity between them is calculated social platform has become an important platform for
by different similarity measurement methods, which is obtaining information and exchanging views in our lives.
recorded as sim(i, j). This paper uses the modified cosine Most of us begin to rely on social media for sharing,
similarity calculation method: communication, discovery and socialization. At the same
time, a large number of users generate content, status
EcI i , j Ri ,c  R R j ,c  Rc  updates, forwarding, replies, etc., which can be obtained on
simi, j  social media. The user clustering method proposed in this
¦ cI i , j
Ri ,c R i ,c  Rc  u ¦cI
2

i, j
R j ,c  Rc 
2
 (1) paper is based on text classification and dynamic analysis of
user behavior. Subsequent relevant experiments have proved
the feasibility of this method. There is a friendly relationship
Ri ,c is the rating of item c by user i, and Rc is the between users in social networks, and the user's historical
behavior data set contains the user's interest data [12]. The
average rating of item c. After calculating the similarity content-based recommendation system is simple and
between users, for a user u, a set of "neighbors" arranged effective, and has the characteristics of low complexity and
according to the similarity is generated, low calculation cost. The recommendation system analyzes
N ^U1 ,U 2 ,},U t ` , 0≤t≤m, u does not belong to N, user behavior, finds correlations between users or items, and
from U1 to Ut, simu , U i 1 d i d t  from large to Small
recommends them to users.

arrangement. The low degree of automation recommended is a

problem caused by the untimely human-computer
The generation of recommendations mainly solves the interaction. When the user performs human-computer
prediction of the target user's interest in unrated items from interaction, he informs the system of his own information
the nearest neighbor information. The prediction of user and products he is interested in. However, due to the lack of
interest can be calculated as follows: timely system updates or delay effects, the resulting
recommendation effect is poor. This method is to mine the
time series relationship in the transaction set. Like related
¦ R  Rm u simu, m
n

m ,i data mining, it is to find a universal rule from the user

Pu ,i Ru  m 1 access log. In the distribution comparison, the relationship
n between "forwarding or not" and "posting tool" in behavior
¦ simu, m data is shown in Figure 1.
m 1 (2)

Ru is the average score of user u on resources, Rm,i is

Rm is the average score of
the score of user m on project I,
user m on resources, and simu, m is the similarity
between users u and m.
Before recommending information, the text in the
recommendation set should be processed first. The key of an
e-commerce recommendation system with good results lies
in how to construct user clusters reasonably, in which the
main work in this process is to extract the "tags" of users.
Because the tag is to analyze the basic information of the
user, so as to obtain the highly refined user characteristic
identification [11]. Content-based recommendation was first Figure 1 Relationship between "forwarding or not" and posting tool in
applied to information retrieval and information filtering. In behavior data
other fields involving data statistics, if data loss occurs, the
possibility of data loss can also be predicted based on The database in the e-commerce system records the
historical data. The main idea of content-based transaction data of each customer, each transaction data

198

Authorized licensed use limited to: ULAKBIM UASL - YILDIZ TEKNIK UNIVERSITESI. Downloaded on May 14,2021 at 18:09:57 UTC from IEEE Xplore. Restrictions apply.
 records the goods purchased by the customer, and each topics. As the number of topics increases, the complexity of
commodity has its category attribute, so that the user's the model increases and the time consumption increases.
preference value for different commodity categories can be The evaluation results of the interest model of collaborative
calculated by using these data and the user's evaluation filtering recommendation system and hybrid
information of the goods. The specific methods are as recommendation system are shown in Figure 2.
follows:

¦ iI u
PI u ,i u P j xi 
PCu , j j 1,2,3}
¦ iI u
P j xi 
(3)

In the formula: PCu , j represents the preference value

of user u for the category, PI u ,i represents the rating value
of user u for product i, I u represents the set of products that
user u has evaluated, and P j xi  represents the
membership degree of product i for category j.
The popularity of recommended products is measured by
the average degree of parameters. For example, Table I Figure 2 Evaluation results of interest model
shows the relationship between the average degree and the
length of the recommendation list under different training Most systems can't effectively use the user's history to
set ratios. find data, only based on the current recommendation
information, which will inevitably lead to the decline of
TABLE I RELATIONSHIP BETWEEN AVERAGE DEGREE AND RECOMMENDED
recommendation persistence. Hybrid recommendation
LIST LENGTH system can realize the mining of users' potential interests in
collaborative filtering algorithm, and recommend items with
Length of Average degree of Average degree of
recommendation collaborative hybrid
few features and low interpretability. The general needs of
list filtering algorithm recommendation users are studied by simple analysis method, and how to
system control resources is determined by analysis. Aiming at the
20 323.2 318.5 subjective and subtle user credit attribute, we try to use
50 314.4 322.4 social data to predict. It is found that both text data and
100 269.7 294.2 behavior data can predict users' credit attributes to a certain
200 279.3 208.1
extent. Depending on the operation of the user set with
The collaborative filtering recommendation system similar current user behavior, it avoids the disadvantage that
relies on the user's actions on items, avoiding the drawback the recommendation system based on content filtering can
that the content filtering-based recommendation system can only produce recommendations similar to the existing
only generate recommendations that are similar to existing interests, and can better dig out the potential interests of
interests, which can better tap the user's potential interests users and have the ability to recommend new information.
and recommend new information Ability. For a new user or
new item with a behavior or operation data semicolon, it is IV. CONCLUSION
difficult to form an effective recommendation. The user With the increasing scale of e-commerce, people pay
browsing process is generally anonymous, so there is less more and more attention to the poor scalability of
information about the user's demographic dimensions, and collaborative filtering recommendation algorithm.
the user's browsing history data is relatively easy to obtain, Recommendation system has become an indispensable part
which makes it possible to establish a user interest model. of e-commerce websites. Accurate, dynamic and
We directly use the user activity statistics provided by the personalized recommendation service can provide users
social platform as the "activity" feature to describe user with valuable shopping reference, thus retaining customers
clustering, the additional active days required to improve and finally achieving the purpose of increasing website sales.
and the elapsed time since the user last logged in. The Most systems can't effectively use the user's history to find
statistics of correlation coefficients of personal information data, only based on the current recommendation information,
characteristics are shown in Table II. which will inevitably lead to the decline of recommendation
persistence. At present, the recommendation service
TABLE II STATISTICS OF CORRELATION COEFFICIENTS OF PERSONAL provided by e-commerce websites is difficult to meet the
INFORMATION CHARACTERISTICS personalized needs of users. As the most successful
Gender Correlation coefficient recommendation algorithm, collaborative filtering algorithm
Age 5.27×10-2 still has some problems such as data sparsity, difficulty in
Education 4.36×10-2 obtaining user scores and authenticity. This paper introduces
Geographical position 4.28×10-2 the development of e-commerce recommendation system
Work 4.99×10-2 and various recommendation technologies, and makes
Activity 5.07×10-2
theoretical analysis and research on the improvement of
From the evaluation results, it can be seen that the introducing user clustering into e-commerce
performance of user clustering is related to the number of recommendation system. By effectively explaining and

199

Authorized licensed use limited to: ULAKBIM UASL - YILDIZ TEKNIK UNIVERSITESI. Downloaded on May 14,2021 at 18:09:57 UTC from IEEE Xplore. Restrictions apply.
 evaluating the results of user mining, we can effectively Rules Mining Algorithm [J]. Modern Electronic Technology, Vol. 39,
recommend the products that users may like individually, No. 15, pp. 133-136, 2016.
and at the same time, we can enable enterprises to provide [6] Yuan Yuan. Research on the implementation technology of
personalized recommendation system in e-commerce [J]. Shanxi
personalized information services to customers and carry Electronic Technology, No. 2, PP. 89-90, 2016.
out targeted e-commerce activities. [7] Yuan Chaofan. Application research of social media in e-commerce
recommendation system [J]. New Media Research, Vol. 3, No. 23, pp.
REFERENCES 14-16, 2017.
[1] Li Chao. Analysis of personalized recommendation system for e- [8] Li Fei. Personalized recommendation system based on optimization
commerce based on big data[J]. Business Economics Research, Vol. of association rules [J]. Journal of Inner Mongolia Normal University
765, No. 02, pp. 71-74, 2019. (Natural Science Chinese Edition), No. 4, pp. 515-520, 2016.
[2] Li Yanjuan, Niu Mengting, Li Linhui. Collaborative filtering [9] Yang Xihui, Lin Peng, Zhou Guoqiang. Collaborative filtering
recommendation algorithm based on bee colony K-means clustering recommendation algorithm based on grey correlation clustering[J].
model [J]. Computer Engineering and Science, Vol. 041, No. 06, pp. Software Guide, Vol. 14, No. 10, pp. 29-34, 2015.
1101-1109, 2019. [10] Yang Guang, Zhong Nai, Xia Hongxia, et al. Collaborative filtering
[3] Wang Xiaojun. Distributed hybrid collaborative filtering method in algorithm of associative clustering based on distributed processing [J].
recommendation system [J]. Journal of Beijing University of Posts Journal of Wuhan University of Technology, Vol. 37, No. 11, pp. 84-
and Telecommunications, Vol. 39, No. 2, pp. 25-29, 2016. 92, 2015.
[4] Dong Hui, Fang Xiao, Ma Jian, et al. Collaborative filtering [11] Li Haixia. Research on personalized recommendation model based on
recommendation algorithm for mobile e-commerce user-item ant colony clustering [J]. Mechanical Design and Manufacturing
clustering based on situational awareness[J]. Journal of Guangxi Engineering, Vol. 46, No. 8, pp. 88-91, 2017.
University for Nationalities (Natural Science Edition), Vol. 024, No. [12] Gao Maoting, Duan Yuanbo. Recommendation algorithm combining
2, pp. 61-68, 2018. user clustering and rating preference [J]. Computer Application
[5] Hao Haitao, Ma Yuanyuan. Research on Electronic Commerce Research, Vol. 035, No.8, pp. 2260-2264, 2018.
Product Recommendation System Based on Weighted Association

200

Authorized licensed use limited to: ULAKBIM UASL - YILDIZ TEKNIK UNIVERSITESI. Downloaded on May 14,2021 at 18:09:57 UTC from IEEE Xplore. Restrictions apply.