Available online at www.sciencedirect.

com
Available online at www.sciencedirect.com

ScienceDirect
ScienceDirect
Available online
Procedia at www.sciencedirect.com
Computer Science 00 (2019) 000–000
Procedia Computer Science 00 (2019) 000–000
www.elsevier.com/locate/procedia
ScienceDirect www.elsevier.com/locate/procedia

Procedia Computer Science 148 (2019) 45–54

Second International Conference on Intelligent Computing in Data Sciences (ICDS 2018)

Second International Conference on Intelligent Computing in Data Sciences (ICDS 2018)

Performance of machine learning techniques in the detection of

Performance of machine learning techniques in the detection of
financial frauds
financial
a frauds
a a
I.SADGALIa, N.SAELa, F.BENABBOUa
I.SADGALI
a
, N.SAEL , F.BENABBOU
Laboratory of Modeling and Information Technology
a
Laboratory
Faculty of sciences of Modeling
Ben M’SIK, and Hassan
University Information Technology
II, Casablanca, Morocco
Faculty of sciences [email protected],
[email protected], M’SIK, University Hassan II, Casablanca, Morocco
[email protected]
[email protected], [email protected],
00212661822838 [email protected]
00212661822838

Abstract
Abstract
Financial fraud presents more and more threat that has serious consequences in the financial sector. As a result, financial institutions
areFinancial
forced tofraud presentsim-prove
continually more andtheir
more threat
fraud that hassystems.
detection serious consequences in several
In recent years, the financial sector.
studies have As a result,
used machinefinancial institutions
learning and data
are forced
mining to continually
techniques im-prove
to provide theirtofraud
solutions detectionInsystems.
this problem. In we
this paper, recent years,a several
propose state of studies have used
art on various fraudmachine learning
techniques, andasdata
as well de-
tection and prevention techniques proposed in the literature such as classification, clustering, and regression. The aim of this studyasisde-
mining techniques to provide solutions to this problem. In this paper, we propose a state of art on various fraud techniques, as well to
tection and
identify the prevention
techniques techniques
and methodsproposed
that giveinthe
thebest
literature
resultssuch as classification,
that have been perfected clustering,
so far. and regression. The aim of this study is to
identify the techniques and methods that give the best results that have been perfected so far.
© 2019
© 2019 The
I.SADGALI, N.SAEL, F.BENABBOU.
Authors. Published by Elsevier B.V. Published by Elsevier B.V.
© 2019
This I.SADGALI,
is an
an accessN.SAEL,
open access F.BENABBOU.
article under
under the CC
CC BY-NC-NDPublished
BY-NC-ND by Elsevier
license B.V.
(http://creativecommons.org/licenses/by-nc-nd/3.0/)
This is open article the license (https://creativecommons.org/licenses/by-nc-nd/4.0/)
This is an open access article under CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/3.0/)
Peer-review under responsibility of the scientific committee of the Second International
Peer-review under responsibility of the scientific committee of the Second InternationalConference
Conferenceon
onIntelligent
IntelligentComputing
Computinginin
Peer-review
Data Sciences
Sciencesunder responsibility
(ICDS 2018). of the scientific committee of the Second International Conference on Intelligent Computing in
Data (ICDS 2018).
Data Sciences (ICDS 2018).
Keywords: Fraud detection, financial fraud, machine-learning, performance;
Keywords: Fraud detection, financial fraud, machine-learning, performance;

1.Introduction
1.Introduction
Financial fraud affects tremendously both the financial industry and everyday life. Fraud can reduce confidence in
theFinancial
industry,fraud affectssavings
destabilize tremendously boththe
and affect thecost
financial industry
of living. and everyday
Financial life.use
institutions Fraud can reduce
a variety confidence
of fraud in
prevention
the industry,
models destabilize
to address savings and
this problem. affect the
However, cost of living.
fraudsters Financial
are adaptive, andinstitutions
over time,use a variety
they conceiveof fraud prevention
several ways of
models tosuch
intruding address this problem.
protective models. However,
Despite thefraudsters areofadaptive,
best effort financial and over time,
institutions, lawthey conceive and
enforcement several ways of
government,
intruding such protective models. Despite the best effort of financial institutions, law enforcement
financial fraud continues to grow. Fraudsters today can be a very inventive, intelligent and fast fraternity. and government,
financial fraud
This paper, continues
seeks to carryto out
grow. Fraudstersanalysis
comparative today canofbe a very inventive,
financial intelligent
fraud detection and fast like
techniques, fraternity.
machine-learning
This paper, seeks to carry out comparative analysis of financial fraud detection techniques,
techniques, who plays an important role in fraud detection, as it is often applied to extract and uncover like machine-learning
the hidden
techniques, who plays an important role in fraud detection, as it is often applied to extract
truths behind very large quantities of data. Also, many modern techniques for detecting fraud are continually and uncover the hidden
truths behind very large quantities of data. Also, many modern techniques for detecting fraud
involves and applied to many areas due to the remarkable lift of fraud which effects on financial field in each year. are continually
involves
Our and applied
objective to many
is to point areas
out their due toand
strength the weaknesses
remarkable andlift of
alsofraud
aimwhich effects
to identify theon financial
open issues field in each
of fraud year.
analysis.
Our objective is to point out their strength and weaknesses and also aim to identify the open issues of fraud analysis.

1877-0509 © 2019 I.SADGALI, N.SAEL, F.BENABBOU. Published by Elsevier B.V.

1877-0509 © 2019
This is an open I.SADGALI,
access N.SAEL,
article under the CC F.BENABBOU.
BY-NC-ND license Published by Elsevier B.V.
(http://creativecommons.org/licenses/by-nc-nd/3.0/)
This is an open
Peer-review access
under article under
responsibility CC BY-NC-ND
of the scientific license
committee (http://creativecommons.org/licenses/by-nc-nd/3.0/)
of the Second International Conference on Intelligent Computing in Data
Peer-review ©under
Sciences (ICDS
1877-0509 responsibility
The Authors. of
2018).
2019 the scientific
Published committee
by Elsevier B.V. of the Second International Conference on Intelligent Computing in Data
Sciences
This (ICDS
is an open2018).
access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the Second International Conference on Intelligent Computing in
Data Sciences (ICDS 2018).
10.1016/j.procs.2019.01.007
46 I. SADGALI et al. / Procedia Computer Science 148 (2019) 45–54
2 I.SADGALI/ Procedia Computer Science 00 (2019) 000–000

The rest of this paper is organized as follows. Section 2 contains a definition and types of financial fraud, Section 3
present techniques implemented for financial fraud detection, Section 4 present review of related work. The
comparative study of various techniques is detailed in Section 5. Finally, we discuss the results and Future work in
Section.

2.Financial fraud
2.1 Definition
Fraud definition, according to the Association of Certified Fraud Examiners (ACFE) “ACFE Association of Fraud
Examiners Certificates”, fraud includes any intentional or deliberate act of depriving another of property or money
by cunning, deception or other unfair acts [12].

2.2 Types of financial fraud

There are several types of financial fraud; we present here a brief description of some of the main types of fraud.
Insurance fraud can occur at many points in the insurance process (e.g., application, eligibility, rating, billing, and
claims), and can be committed by consumers, agents and brokers, insurance company employees, healthcare
providers, and others [1, 2]
Securities and commodities fraud, the FBI [3] provides brief descriptions of some of the most prevalent securities
and commodities frauds encountered today, for example, “Market Manipulation, High Yield Investment Fraud, The
Ponzi Scheme, The Pyramid Scheme, Prime Bank Scheme, Advance Fee Fraud, Hedge Fund Fraud, Commodities
Fraud, Foreign Exchange Fraud, Broker Embezzlement and Late-Day Trading.” According to another definition by
CULS [4], securities frauds include theft from manipulation of the market, theft from securities accounts, and wire
fraud.
Money Laundering is the process by which criminals conceal or disguise the proceeds of their crimes or convert
those proceeds into goods and services. It allows criminals to inject their illegal money into the stream of commerce,
thus corrupting financial institutions and the money supply and giving criminals unwarranted economic power [5].
Gao and Ye [6] similarly define money laundering as the process by which criminals “wash dirty money” to
disguise its illicit origin and make it appear legitimate and “clean.”
Financial statement fraud (corporate fraud), financial statements are a company's basic documents to reflect its
financial status [10]. It had an objective as.
• Fraud these statements to make the business more profitable
• Improvement of the performance of the actions
• Reduction of tax obligations
• Attempt to exaggerate performance due to managerial pressure
Credit card fraud is essentially of two types; application and behavioural fraud [26]. Application fraud is where
fraudsters obtains new cards from issuing companies using false information or other people's information.
Behavioural fraud can be of four types: mail theft, stolen/lost card, counterfeit card and ‘card holder not present’
fraud [25]
Mortgage Fraud is a specific form of financial fraud that refers to the manipulation of a property or mortgage
documents. It is often committed to distort the value of a property for the purpose of influencing a lender to finance
a loan for it [5].

3. Financial Fraud detection techniques

As stated above, we investigated techniques used for financial fraud detection in anterior works, this bellow our
suggested classification for this techniques and a small description for each one of them.

3.1 Descriptive or Unsupervised Techniques

This paragraph, describes descriptive models, that is, the unsupervised learning functions. These functions do not
predict a target value, but focus more on the intrinsic structure, relations, interconnectedness, etc.
 I. SADGALI et al. / Procedia Computer Science 148 (2019) 45–54 47
I.SADGALI/ Procedia Computer Science 00 (2019) 000–000 3

Self-Organizing Maps [30] a self-organizing map (SOM) is a neural network technique but used unsupervised
learning. SOM allows users to visualize data from high dimensional to low dimensional.
Group method of data handling (GMDH) is an inductive learning algorithm for modeling complex systems. It is a
self-organizing approach that tests increasingly complicated models and evaluates them using some external
criterion on separate parts of the data sample [11]
Outlier detection methods (OD) is very different from the traditional observation method. Outlier method is used
to detect unusual behavior of a system using a different mechanism. [38].
Association rule analysis (AR) are defined on transaction sets. Given that it is more common to work with tuples
rather than transactions in a database, various solutions to this problem has been proposed. When working with
relational databases, it is usual is to consider each item to be a pair (attribute, value) and each transaction to be a
tuple in a table [22].
Density based spatial clustering of applications with noise (DBSCAN) is a density based clustering algorithm
which can be used to filter out outliers and discover clusters of arbitrary shapes [23].

3.2 Predictive Techniques

In predictive analysis, the purpose is to build an analytic model that predicts target objects of interest.
Logistic Regression (LR) logistic regression is a type of generalized linear model. Using simple linear regression
is inappropriate when the variable to be predicted is binary; due to normality assumptions. [7].
Decision Trees (DT) is a tree structure, where each node represents a test on an attribute and each branch represents an
outcome of the test. In this way, the tree attempts to divide observations into mutually exclusive subgroups [14].
Classification and Regression Tree (CART) is a computerized, non-parametric technique different from traditional
statistical methods. CART applies the binary Recursive Partitioning Algorithm (RPA) to best classify samples into a
number of non-overlapping regions, each of which corresponds to a terminal node in the tree [20].
Decision Trees C4.5 gives algorithm and solutions to a set of problems that have arisen over the years among
decision tree researchers like handling various problems such as missing attribute values [19].
Cost-sensitive decision tree (CSDT) an induction algorithm developed to identify fraudulent credit card
transactions are given. In the well-known decision tree algorithms, the splitting criteria are either insensitive to costs
and class distributions or the cost is fixed to a constant ratio [29].
Neural Networks (NN) is a mature technology with an established theory and recognized application areas. A
NN consists of a number of neurons, i.e., interconnected processing units. Associated with each connection is a
numerical value, called "weight"[14].
Probabilistic neural network (PNN) is a feed-forward NN involving a one pass training algorithm used for
classification and mapping of data. It is a pattern classification network, based on the classical Bayes classifier,
which is statistically an optimal classifier that seeks to minimize the risk of misclassification [11].
Support Vector Machines (SVM) use a linear model to implement nonlinear class boundaries by mapping input
vectors nonlinearly into a high-dimensional feature space. In the new space, an optimal separating hyperplane is
constructed [11].
Naïve Bayes (NB) a classification tool simply uses Bayes conditional probability rule. Each attribute and class
label are considered random variable, and assuming that the attributes are independent, the naïve Bayes finds a class
to the new observation that maximizes its probability given the values of the attributes.[9].
Bayesian belief network (BBN) allow for the representation of dependencies among subsets of attributes. A BBN
is a directed acyclic graph, where each node represents an attribute and each arrow represents a probabilistic
dependence [14].
Bayesian skewed logit model (BSL) this model incorporates the possibility of using asymmetric links in order to
measure the probability of yi = 0 and yi = 1 in non-balanced samples [9]
K-nearest neighbor (KNN) is used largely in detection systems. It is also proved that KNN works extremely well
in credit card fraud detection systems using supervised learning techniques. [38].
Bivariate Probit Model (BP) is typically used where a dichotomous indicator is the outcome of interest and the
determinants of the probable outcome includes qualitative information in the form of a dummy variable where, even
after controlling for a set of covariates, the possibility that the dummy explanatory variable is endogenous cannot be
ruled out a priori [40].
48 I. SADGALI et al. / Procedia Computer Science 148 (2019) 45–54
4 I.SADGALI/ Procedia Computer Science 00 (2019) 000–000

3.3 Artificial & Computational Intelligence Techniques

This part, describes artificial and computational intelligence models, which is, a set of nature-inspired
computational methodologies and approaches to address complex real-world problems to which mathematical or
traditional modelling.
Genetic Algorithm (GA) in Genetic Algorithm i.e. inspired from natural evolution, randomly generated rules are
considered as an initial population[15].
Genetic programming (GP) is an extension of genetic algorithms (GA). It is a search methodology belonging to
the family of evolutionary computation. GP randomly generates an initial population of solutions. Then, the initial
population is manipulated using various genetic operators to produce new populations [11].
Scatter Search (SS) is an evolutionary algorithm, which shares some common characteristics with the GA. It
operates on a set of solutions, the reference set, by combining these solutions to create new ones [27].
Hidden Markov Model (HMM) it differs from the normal statistical Markov model by having invisible states, but
each state randomly generates one of the visible states. A hidden Markov model can be presented as the simplest
dynamic Bayesian network [36].
Iterative Dichotomiser 3 (ID3) for dealing with symbolic data by expressing the knowledge as a decision tree [39].
Artificial Immune System (AIS) the human biological immune system has a number of fundamental characteristics
that can be adapted as design principles for AIS applications in various problem domains [28].
Artificial Immune Recognition System (AIRS) both self/non-self cells and detector cells are represented as feature
vectors. In order to reduce redundancy, ARB (Artificial Recognition Ball) is used which is representative of similar
memory cells [31].
Artificial neural network (ANN) artificial neural networks were first created with the purpose to imitate the behavior of
the human brain. A neural network is the connection of elementary objects called the simple neuron [32].
Multilayer Perception Algorithm (MPL) is an artificial neural network and is a nonparametric estimator that can
be using for classifying and detecting intrusions [32].
Parenclitic Network (PN) a network reconstruction technique that allows highlighting the differences between one
instance and a set of standard [33].
Multi-layer feed forward neural network (MLFF-NN) is one of the most common NN structures, as they are
simple and effective, and have found home in a wide assortment of machine learning applications [11].

3.4 Other concepts

This paragraph, present some concepts, which associated with techniques above, for hybrids models.
Fuzzy logic for representing the cognitive uncertainties, measuring the intensity of the truth values for
unquantifiable measures or probabilistic measures within the range of 0 and 1 [39]. Fuzzy association rules (FAR)
can be found in the literature such as a generalization of association rules when initial data are fuzzy or if they have
been previously processed to provide them with imprecision [22].
Dempster Shafer Theory (DST) or evidence theory is a general framework for reasoning with uncertainty, the
role of DST is to combine evidences from the rules R1and R2 and compute an overall belief value for each
transaction. [23].
Computational fraud detection model (CFDM) using SAS® Enterprise Miner™ (EM) as an automation tool to
develop the model. The process retains maximal information and uses essentially all of it in processing the
documents. [18].
Locally Weighted Learning (LWL) represent nonlinear functions, yet has simple training rules with a single global
optimum for building a local model in response to a query. This allows complex nonlinear models to be identified
(trained) quickly [41].

4 Related Work
For each type of fraud, several techniques has been used each of which has advantages and shortcomings.

4.1 Insurance fraud

 I. SADGALI et al. / Procedia Computer Science 148 (2019) 45–54 49
I.SADGALI/ Procedia Computer Science 00 (2019) 000–000 5

In 2007, StijnViaene found that with claim amount information available at screening time detection rules can be
accommodated to increase expected profits, he used logistic regression model [7]. Jean Pinquet presented a
statistical approach that counteracts selection bias without using a random auditing strategy [8]. In 2008, the use of
an asymmetric link notably improves the percentage of cases that are correctly classified after the model estimation
[9], showed by Bermudez.

4.2 card fraud

In 2008, Quah and Sriganesh. proposed, for real-time fraud detection, a new and innovative approach; it makes
use of self-organization map, Neural Networks and rules induction [21].
In 2009, a novel methodology has been applied, on credit card fraud, using AR and FAR, was proposed by Sanchez
[22]. Suvasini investigated a fusion approach using Dempster–Shafer theory and Bayesian learning [23]. Whitrow
developed a framework for transaction aggregation, using a variety of classification methods and a realistic cost-
based performance measure [24].
In 2011, Bhattacharyya evaluated two advanced data mining approaches, support vector machines and random
forests, together with the well-known logistic regression [25]. Duman suggested a novel combination of the two
well-known meta-heuristic approaches, namely the genetic algorithms and the scatter search. The method was
applied to real data and very successful results were obtained [27].
In 2012, Wong proposed the use of AIS on one aspect of security management; the detection of credit card fraud
[28]. In 2013, Sahin developed a methodology for fraud detection using decision tree and showed that this cost-
sensitive decision tree algorithm outperforms the existing methods [29].
In 2014, Olszewski proposed a method of the detection threshold setting based on SOM [30]. Halvaiee developed
a novel model for credit card fraud detection using AIS and introduced a new model called AIS-based Fraud
Detection Model (AFDM), increase the accuracy up to 25%, reduce the cost up to 85%, and decrease system
response time up to 40% comparedto the base algorithm [31].
In 2015, L.Dhanabal analysed NSL-KDD data set, and used it to study the effectiveness of the various
classification algorithms in detecting the anomalies [35], the analysis was done using classification algorithms
available in the data-mining tool WEKA.
In 2016, Dai developed a hybrid framework with Big Data technologies [36]; that implement it with latest big data
technologies like Hadoop, Spark, Storm, HBase, which showed great potentials of achieving the goals. Adewumi focused
on recent Machine Learning based and Nature Inspired based credit card fraud detection techniques proposed
in literature [37].
In 2017, Mubalaik proposed an ANN-MPL multilayer perception [32], based on implementation of well-known
machines learning techniques, it helps to anticipate and quickly detect fraud. Zanin proposed hybrid data mining /
complex network, classification algorithm, able to detect illegal instances in a real card transaction data set [33].
Malini implemented KNN algorithm and outlier detection methods to optimize the best solution for the fraud
detection problem [38] These approaches are proved to minimize the false alarm rates and increase the fraud
detection rate. Askari proposed fraud detection algorithm based on Fuzzy-ID3 [39]. Experimental result exhibits that
the technique is efficient one in detecting frauds.

4.3 Financial statement fraud

In 2007, Kirkos investigated the usefulness of Decision Trees, Neural Networks and Bayesian Belief Networks in
the identification of fraudulent financial statements [14].Genetic algorithm approach was proposed by HOOGS the
patterns are capable of identifying potentially fraudulent behavior despite occasional missing values, and provide
low false positive rates [15].
In 2008, BAI proposed in [20] Classification and Regression Tree (CART), to identify and predict the impacts of
False Financial Statements (FFS).
In 2011, Cecchini developed a methodology for automating ontology creation using WordNet [16]. Humpherys
proposed a parsimonious model with Naïve Bayes and C4.5 achieved the highest classification accuracy [17] and Glancy
proposed, for detecting fraud in financial reporting, a computational fraud detection model, using a quantitative
 50 I. SADGALI et al. / Procedia Computer Science 148 (2019) 45–54
6 I.SADGALI/ Procedia Computer Science 00 (2019) 000–000

approach
6 on textual data [18]. Also, Ravisankar
I.SADGALI/ Procediagave a comparison
Computer of 000–000
Science 00 (2019) data mining techniques; Multilayer Feed
Forward Neural Network (MLFF), Support Vector Machines (SVM), Genetic Programming (GP), Group Method of
Dataapproach
Handlingon(GMDH), Logistic
textual data Regression
[18]. Also, (LR),gave
Ravisankar and aProbabilistic
comparison Neural
of data Network (PNN) [11]
mining techniques; in the same
Multilayer Feedyear.
Forward Neural Network (MLFF), Support Vector Machines (SVM), Genetic Programming (GP), Group Method of
Data Handling (GMDH), Logistic Regression (LR), and Probabilistic Neural Network (PNN) [11] in the same year.
5 A comparative study

In5 this
A comparative study
section, we will analyze the contribution of each technique and its effectiveness, in order to find a
promising combination for future work.
In this section, we will analyze the contribution of each technique and its effectiveness, in order to find a
promising combination for future work.
5.1 Criteria
5.1 Criteria
In our comparative tables, we regroup and synthesize most used criteria in anterior works in order to have most
complete comparison:
In our comparative tables, we regroup and synthesize most used criteria in anterior works in order to have most
complete comparison:
• Real time: parameter show if the technique is able to run in real time (R) or not (NR).
• Real time: parameterparameter
show if theof
technique is able to run in real time (R) or not* (NR).
• Accuracy: a validation precision (TP+TN)/ (TP+FP+TN+FN) *
• Accuracy:
• Sensitivity (orarecall):
validation parameter
is the measureof of
precision (TP+TN)/of(TP+FP+TN+FN)
the proportion the number of fraudulent cases predicted correctly as
• Sensitivity
fraudulent (or recall):model,
by a particular is the measure of the
gives the proportion
accuracy of the
on the number
fraud casesofTP/
fraudulent
(TP+FP)cases predicted correctly as
fraudulent by a particular model, gives the accuracy on the
• Dataset: size, type (particular (P), standard (S) or generic (G)) fraud cases TP/ (TP+FP)
• Dataset: size, type (particular (P), standard (S) or generic (G))
• Observations: limitations (-) and contributions (+) of the technique
• Observations: limitations (-) and contributions (+) of the technique

5.2 5.2
Comparative tables
Comparative tables
In this paragraph,
In this wewe
paragraph, present
presentthree
threetables,
tables,for
for each
each type offraud,
type of fraud,we
wegive
givea summary
a summary of techniques
of techniques in past
in past workwork
and and
observations belong
observations thisthis
belong study.
study.

Table
Table 1. 1. Contributionsand
Contributions andlimitations
limitations of
ofML
MLtechniques
techniquesapplied to Insurance
applied frauds
to Insurance frauds
Validation
Validation Data set
Data set
Observations
Observations
Technique
Technique

Sensitivity

Reference
Real time

(%)Accuracy

Sensitivity

Reference
Real time

Accuracy

Type
TP (%)

Size

Type
TP (%)

Size
(%)
(%)

(%)

LR NA 99.42 66,67 NA Claims during 2000 P - The expected cost can be unprofitable [7]
LR NA 99.42 66,67 NA Claims during 2000 P for- the
Thecompany
expected cost can be unprofitable [7]
for the company
BP NA NA NA NA Claims during 2000 P + The expected overestimation of fraud risk derived [8]
BP NA NA NA NA Claims during 2000 P was corrected.
+ The expected overestimation of fraud risk derived [8]
- Present a lack of fit due to the incorrect
was corrected.
10 000 automobile claims P classification of zero cases [9]
BSL NA 99,5 98,46* NA
- Present a lack of fit due to the incorrect
- Unable to signal the significance of the parameter
10 000 automobile claims P classification
associated to theofvariable
zero cases
proxim [9]
BSL NA 99,5 98,46* NA - Unable to signal the significance of the parameter
associated to the variable proxim

*
TP: true positive / TN: true negative / FP: False positive / FN: False negative

*
TP: true positive / TN: true negative / FP: False positive / FN: False negative
 I.SADGALI/ Procedia Computer Science 00 (2019) 000–000 7
I. SADGALI et al. / Procedia Computer Science 148 (2019) 45–54 51

Table 2. Contributions and limitations of ML techniques applied to credit card frauds †

Validation Data set

Observations

Sensitivity (%)
Technique

Real time

Reference
Accuracy

Type
TP (%)

Size
(%)

SOM over 200 million + Clustering helps in identifying new hidden [21]
+NN + R NA NA NA customers P patterns in input data
RI +the filtering of transactions for further review reduces the
overall cost as well as processing time.
12,107 +The applied methodology overcomes the difficulties of
AR + R NA NA NA transactions P minimum support and confidence, optimizes the execution
FAR times, reduces the excessive generation of rules, and helps [22]
make the results more intuitive, thereby facilitating the
work of fraud analysts
DST + R NA 98 NA NA G + architecture has been kept flexible so that new rules using [23]
NB any other effective technique can also be included at a later
stage
RF + AG NR NA NA NA 175million
SVM + NR NA NA NA transactions
AG + The aggregation period has a major impact upon the
LR + AG NR NA NA NA 1.1million P performance of classifiers for fraud detection. [24]
KNN+ NR NA NA NA transactions
AG
SVM NR 95,30 NA 72,70 2420 fraudulent G + While sensitivity, and accuracy decreased with lower [25]
RF NR 90,80 NA 52,40 transactions proportions of fraud in the training data, precision showed
LR NR 94,20 NA 65,40 an opposite trend
GA + SS NR NA NA NA 100,000 P + Bank Management should increase the monitoring [27]
fraudulent capacity if they want to face less losses due to fraud
+ three mechanisms (new transaction representation and
640 361 total variable width r-contiguous bit matching algorithm,
AIS NR 80 88* NA transactions P vaccination process and memory cell evolution process) [28]
significantly improve the performance of the AIS
SVM NR NA 90.0 NA 978 fraudulent +We cannot use misclassification cost without incorporating
CART NR NA 83,1 NA records the class distribution or an impurity measure in cost
NR NA 92.1 NA 22 million normal P calculations. [29]
CSDT transactions
10,000 accounts + high-dimensional data projected onto a 2-dimensional
of selected credit space can be easily analyzed and interpreted even by a non-
SOM NR 100 NA NA card (1.01.2005 - P expert. [30]
1.03.2005)
AIRS + 3.74% fraudulent P +improving memory cell generation improves the detection [31]
CC‡ NR NA 83* NA transactions rate.
+Changing distance function, performs better regarding FP.
DT NR 91,03 NA NA S + Reduce the complexity of host based analysis engines.
NB NR 99,02 NA NA NSL-KDD dataset - Tend to rely on the innate logging and monitoring [32]
ANN+ NR 99,47 NA NA [34] capabilities of the server.

† NA: Not Addressed

Proxim: Accident occurred between the policy issue date and the effective starting date, 1; otherwise 0.
*: Calculated
‡ CC: cloud computing
52 I. SADGALI et al. / Procedia Computer Science 148 (2019) 45–54
8 I.SADGALI/ Procedia Computer Science 00 (2019) 000–000

MPL
PN+ Transactions 01- P +the addition of parenclitic features to the raw data set [33]
→
ANN + NR NA NA NA 2011 12- 2012. enhance the obtained results, with the error dropping from
MPL a 19:2 to a 12:23%.
SVM NR 98,8 NA NA +NSL-KDD dataset show that it is a best candidate data set
to simulate and test the performance of IDS.
NB NR 74,9 NA NA NSL-KDD S +Correlation, based Feature Selection method for [35]
dimensionality reduction, reduces the detection time and
increase the accuracy rate.
DBSCAN 5 dataset, -should allocate ore computing resources to streaming
+ HMM 10,000,.. Detection Layer and Batch Training
+ LR R NA NA NA 1,000,000 within G Layer to improve the overall performance during real [36]
one year. system deployment
KNN+ +OD works fast and well on online large datasets [38]
OD NR NA NA NA NA P +KNN can suit for detecting fraud with the limitation of
memory
FL+ ID3 NR 89 NA NA NA P +reduce the irrelevant processing so that the detection can [39]
be done in optimal time

Table 3. Contributions and limitations of ML techniques applied to financial statement frauds §

Validation Data set

Observations
Accuracy (%)

Sensitivity

Reference
Technique

Real time

TP (%)

Type
Size
(%)

DT NA 73.6 75 NA P + associated falsification with financial distress, since it used Z [14]

NN NA 80 82.5 NA 76 Greek firms score as a first level splitter
BBN NA 90.3 88.9 NA + revealed dependencies between falsification and the ratios debt
to equity, net profit to total assets, sales to total assets, working
capital to total assets and Z score.
AAERs published S + a successful technique for detecting discriminatory patterns in
by the SEC challenging domains characterized by high dimensionality
GA NA 95 63 NA between 2002- + Patterns are easily translated to domain appropriate language [15]
→
2004 easily understood by external stakeholders.
Ontolo MDAs for 78 S +The methodology can be applied to available
gy+ NA 83.87 81.97 NA companies text for any financial problem where the goal is to create a [16]
WN** between 1994 to dictionary
1999 (ontology) of discriminating concepts.
LR NA 63.4 62,9 NA S +The four-variable model has the predictive accuracy equal to
NB NA 67.3 66.7 NA the 10-variable model and performs better than the 24-variable
SVM NA 65.8 64.3 NA 101 companies model. [17]
C4.5 NA 67.3 68.0 NA +The best performance came from the Naïve Bayes classifier and
LWL NA 60.4 60.6 NA the C4.5 decision tree classifiers using the 10-variable model.
AAERs published S + has the potential to serve as a filtering tool for regulators to [18]
CFDM NA 90,9* 80* NA by the SEC 2006- focus their resources and subsequently increase the detection of
2008 financial reporting fraud
MLFF- NA 78.36 NA 80.21
NN
SVM NA 70.41 NA 55.43 202 companies: + Hybrid data mining techniques that combine two or more

§ AAER: Accounting and Auditing Enforcement

Release SEC: Securities and Exchange Commission
MDAs: Management's Discussion and Analysis
AG: Aggregation
** WN: WordNet
 I. SADGALI et al. / Procedia Computer Science 148 (2019) 45–54 53
I.SADGALI/
I.SADGALI/ Procedia ComputerScience
Procedia Computer Science
0000 (2019)
(2019) 000–000
000–000 9 9

GP GP NA NA 94.14
94.14 NANA 95.09
95.09 101
101fraudulent 101
fraudulent 101 classifiers
classifiers cancan
be be used
used on the
on the samesame dataset.
dataset.
GMDHGMDHNA NA 93.00
93.00 NANA 91.46
91.46 non-fraudulent
non-fraudulent ++Results
Results areare
much
muchsuperior to antoearlier
superior studystudy
an earlier on theon
same
the same
LR LR NA NA 66.86
66.86 NANA 63.32
63.32 dataset.
dataset.
PNN PNN NA NA 98.09
98.09 NANA 98.09
98.09 SS [11] [11]
CARTCARTNA NA 72.38
72.38 NANA 72.40
72.40
CART NA NA 98,39 NA 24 false, 124 non- S + CART produce more accurate classification on the fraud cases [20]
CART NA NA 98,39 NA 24 false, 124 non-
false: financial
S + CART produce more accurate classification on the fraud cases [20]
false: financial
LR NA NA 95,97 NA reports
LR NA NA 95,97 NA reports
5.3 Synthesis and discussion
5.3 Synthesis and discussion
It could be observed that almost, all implemented algorithms, do not work in real time.
It could
As can be observed
be seen, the thatdetection
almost, allof implemented
credit card fraud algorithms, do not
uses several MLwork in real time.
techniques, especially those of artificial
Asintelligences
can be seen, and the detection
combines themofwithcredit card fraud
optimization uses several
techniques such asML techniques,
aggregation, especially
for the detectionthose of artificial
of frauds of
intelligences and combines
financial statements themmainly
it is based with onoptimization
text processingtechniques
techniques.such as aggregation, for the detection of frauds of
financialFor statements it is based
fraud insurance, mainly on textofprocessing
the non-necessity the real-time techniques.
processing, makes the detection of the fraud easier,
nevertheless
For the difficulty
fraud insurance, the resides in the fact
non-necessity of that
the these deceptions
real-time are human
processing, and can
makes the bedetection
well masked.
of theComparing
fraud easier,
the logistic regression and the Bayesian results, we see that the Bayes logistic model
nevertheless the difficulty resides in the fact that these deceptions are human and can be well masked. Comparing gives posterior estimations for
the true positive
the logistic regressionrate.and the Bayesian results, we see that the Bayes logistic model gives posterior estimations for
In financial statement fraud, results based on the accuracy indicated that the PNN was the best performing
the true positive rate.
(98.09%) following by Genetic algorithm (95%) who gave marginally lower accuracies in most cases.
InNaives
financial statement fraud, results based on the accuracy indicated that the PNN was the best performing
bays and SVM gives good results with NSL-KDD dataset (99,02%, 98,8%) for credit card fraud. Also we
(98.09%) following
found that: by Genetic algorithm (95%) who gave marginally lower accuracies in most cases.
Naives bays and SVM
− The aggregation period gives hasgood results
a major with
impact NSL-KDD
on the performance dataset (99,02%,
for fraud 98,8%)
detection. for credita product
Aggregating card fraud. Also we
foundimproves
that: the prediction rate for all techniques except for CART.
− The aggregation
− SOM period
Clustering helpshas to aidentify
major newimpact on the
patterns performance
hidden for data,
in the input fraudwhich
detection. Aggregating
otherwise cannot beaidentified
product by
improves the prediction
traditional statistical rate for alltransaction
methods, techniquesfiltering
except for for CART.
further examination reduces overall cost as well as
− SOM processing
Clusteringtime.
helps to identify new patterns hidden in the input data, which otherwise cannot be identified by
− Cost-sensitive
traditional decision
statistical tree approaches
methods, is used
transaction in creditfor
filtering cardfurther
fraud detection
examinationand show that it outperforms
reduces overall costtheas well as
models
processing time.built using the traditional data mining methods such as decision trees, ANN and SVM
− Logistic Regression
− Cost-sensitive decisionworks well with linear
tree approaches datainforcredit
is used creditcard
cardfraud
fraud detection
detection. and show that it outperforms the
− Support
models builtvector
usingmachine method data
the traditional is capable
miningof detecting
methods the such fraudulent activity
as decision at the
trees, ANN time
andof SVM
transaction.
− Complex
− Logistic networks works
Regression can be usedwell as a way
with to improve
linear data for data mining
credit cardmodels;
fraudthey may be integrated as complementary tools
detection.
in a synergistic manner in order to improve the classification rates obtained by classical data mining algorithms.
− Support vector machine method is capable of detecting the fraudulent activity at the time of transaction.
− KNN method can suit for detecting fraud with the limitation of memory. By the meantime, outlier detection
− Complex networks
mechanism helpscan
to be usedtheascredit
detect a waycard
to improve dataless
fraud using mining
memorymodels; they may berequirements.
and computation integrated as complementary tools
in −a Outlier
synergistic manner
detection in order
works to improve
fast and the classification
well on online large datasets. rates obtained by classical data mining algorithms.
− KNN method can suit for detecting fraud with the limitation of memory. By the meantime, outlier detection
mechanism helps to detect the credit card fraud using less memory and computation requirements.
6 Conclusion
− Outlier detectionand future
works fastwork
and well on online large datasets.
In this study, it was found that hybrid fraud detection techniques are the most used, as they combine the strengths
of several traditional detection methods. In addition, we discover that the studies do not smother all types of fraud,
and each type
6 Conclusion of future
and fraud has constraints specific to it; response required in real time, text analysis…
work
In our future work we will
In this study, it was found thatfocushybrid
on continuing the studytechniques
fraud detection of credit card
arefraud to improve
the most used, current
as theyalgorithms,
combine thewe strengths
wish
to include a hybrid model that is both able to handle imbalanced dataset and the real-time problem, to have a
of several traditional detection methods. In addition, we discover that the studies do not smother all types of fraud,
response during the financial transaction runtime, with an improved accuracy.
and each type of fraud has constraints specific to it; response required in real time, text analysis…
In our future work we will focus on continuing the study of credit card fraud to improve current algorithms, we wish
References
to include a hybrid
Coalition against model
Insurance Fraud,that is about
"Learn bothfraud",
ablehttp:
to //www.insurancefraud.org/learn_about_fraud.htm.
handle imbalanced dataset and the real-time problem, to have a
response during the
2. J.L. Kaminski, financial
Insurance Fraud,transaction
OLR Research runtime,
Report, http:with an improved accuracy.
//www.cga.ct.gov/2005/ rpt/2005-R-0025.htm. 2004
3. FBI, Federal Bureau of Investigation, Financial Crimes Report to the Public Fiscal Year, Department of Justice, United States, http:
//www.fbi.gov/publications/financial/fcs_report2007/financial_crime_2007.htm.( 2007)
References
4. CULS, Cornell University Law School, White-Collar Crime: an overview, http: // topics.law.cornell.edu/wex/White-collar_crime (2009)
5. Ngai
Coalition E, HuInsurance
against Y, Wong Y, Chen "Learn
Fraud, Y, and Sun X The
about application
fraud", of data mining techniques in financial fraud detection: A classification
http: //www.insurancefraud.org/learn_about_fraud.htm.
framework
2. J.L. Kaminski,andInsurance
an academic review
Fraud, OLRof literature
ResearchDecision
Report, Support Systems Volume 50, Issue
http: //www.cga.ct.gov/2005/ 3, p559-569 (2011) 2004
rpt/2005-R-0025.htm.
6. Z.Gao,
3. FBI, FederalM.Ye, A framework
Bureau for dataFinancial
of Investigation, mining-based anti-money
Crimes Report laundering research,
to the Public FiscalJournal of Money Laundering
Year, Department Control
of Justice, United10States,
(2), p170–179
http:
//www.fbi.gov/publications/financial/fcs_report2007/financial_crime_2007.htm.( 2007)
4. CULS, Cornell University Law School, White-Collar Crime: an overview, http: // topics.law.cornell.edu/wex/White-collar_crime (2009)
5. Ngai E, Hu Y, Wong Y, Chen Y, and Sun X The application of data mining techniques in financial fraud detection: A classification
framework and an academic review of literature Decision Support Systems Volume 50, Issue 3, p559-569 (2011)
6. Z.Gao, M.Ye, A framework for data mining-based anti-money laundering research, Journal of Money Laundering Control 10 (2), p170–179
54 I. SADGALI et al. / Procedia Computer Science 148 (2019) 45–54
10 I.SADGALI/ Procedia Computer Science 00 (2019) 000–000

(2007)
7. StijnViaene a , Mercedes Ayuso b , Montserrat Guillen b,*, Dirk Van Gheel c , Guido Dedene, Strategies for detecting fraudulent claims in
the automobile insurance industry, Elsevier, European Journal of Operational Research Volume 176, Issue 1, p565-583 (2007)
8. Jean Pinquet Mercedes Ayuso Montserrat Guill´en, Selection bias and auditing policies for insurance claims, The Journal of Risk
and Insurance, Vol. 74, No. 2, p425-440( 2007)
9. Ll. Bermudez, J.M. Perez, M. Ayusoc , E. Gomez , F.J. Vazquez, A Bayesian dichotomous model with asymmetric link for fraud in
insurance, Elsevier, p779- 786(2007)
10. W.H. Beaver, Financial ratios as predictors of failure, Journal of Accounting Research 4 p71–111. (1966)
11. Ravisankar P, Ravi V, Raghava Rao G, and Bose, Detection of financial statement fraud and feature selection using data mining
techniques, Elsevier, Decision Support Systems Volume 50, Issue 2, p491-500 (2011)
12. (Date last accessed 15-July-2014). Online Available: http://www.acfe.com/uploadedfiles/acfewebsite/content/documents/rttn-2010.pdf.
13. D.Zhang, L.Zhou, Discovering Golden Nuggets: Data Mining in Financial Application, IEEE Transactions on Systems, Man,
and Cybernetics, Part C (Applications and Reviews) ( Volume: 34, Issue: 4), p513-522 (2004)
14. E.Kirkos, C.Spathis, Y.Manolopoulos, Data Mining techniques for the detection of fraudulent financial statements, Elsevier, Expert
Systems with Applications Volume 32, Issue 4, p995- 1003(2007)
15. B.Hoogs, T.Kiehl, C.Lacomb, D.Senturk, A genetic algorithm approach to detecting temporal patterns indicative of financial statement,
InterScience, Intelligent systems in accounting, finance and mamagement, Volume15, Issue1‐2, p41-56 (2007)
16. M.Cecchini, H.Aytug, G.Koehler, P.Pathak, Making words work: Using financial text as a predictor of financial events, Elsevier,
Decision Support Systems Volume 50, Issue 1, p164-175 (2010)
17. L. Humpherys, C. Moffitt, B. Burns, K. Burgoon, F. Felix, Identification of fraudulent financial statements using linguistic
credibility analysis, Elsevier, Decision Support Systems Volume 50, Issue3, p585-594.(2010)
18. H. Glancy, B. Yadav, A computational model for financial reporting fraud detection, Elesevier, p596-601(2010)
19. R. Quinlan, C4.5: programs for machine learning, Morgan Kaufmann Publishers, Machine Learning, Volume 16, Issue 3, p235-240 (1994)
20. B.Bai, J.Yen, X.Yang, False financial statements: characteristics of china’s listed companies and cart detection approach,
International Journal of Information Technology & Decision Making Vol. 7, No. 2, p 339-359 (2008)
21. J.Quah, M.Sriganesh, Real-time credit card fraud detection using computational intelligence, Elsevier, Expert Systems with
Applications, Volume 35, Issue 4, p1721-1732 (2008)
22. D.Sa´nchez, M.A. Vila, L. Cerda, J.M. Serrano, Association rules applied to credit card fraud detection, Elsevier, Expert Systems with
Applications, Volume 36, Issue 2, Part 2, p3630-3640(2009)
23. S.Panigrahi,A.Kundu, S.Sural, A.K.Majumdar, Credit card fraud detection: A fusion approach using Dempster–Shafer theory and Bayesian
learning, Elsevier, Information Fusion, Volume 10, Issue 4, p354-363 (2009)
24. C. Whitrow, D. J. Hand, P. Juszczak, D. Weston, N. M. Adams, Transaction aggregation as a strategy for credit card fraud
detection, Springer, Data Mining and Knowledge Discovery, Volume 18, Issue 1, p30-55 (2009)
25. S.Bhattacharyya, S.Jha, K.Tharakunnel, J.C. Westland, Data mining for credit card fraud: A comparative study, Elsevier, Decision
Support Systems, Volume 50, Issue 3, p602-613(2011)
26. Y. Jin, R.M. Rejesus, B.B. Little, Binary choice models for rare events data: a crop insurance fraud application, Applied Economics Volume
37, Issue 7, p841–848. (2005)
27. E.Duman, H.Ozcelik, Detecting credit card fraud by genetic algorithm and scatter search, Elsevier, Expert Systems with
Applications, Volume 38, Issue 10, p13057-13063(2011)
28. N.Wong, P.Ray, G.Stephens, L.Lewis, Artificial immune systems for the detection of credit card fraud: an architecture, prototype
and preliminary results, Information Systems Journal, Volume22, Issue1, p53-76(2012)
29. Y.Sahin, S.Bulkan, E.Duman, A cost-sensitive decision tree approach for fraud detection, Elsevier, Expert Systems with
Applications, Volume 40, Issue 15, p5916-5924 (2013)
30. D.Olszewski, Fraud detection using self-organizing map visualizing the user profiles, Elsevier, Knowledge-Based Systems, Volume 70, p324-
333 (2014)
31. N.S.Halvaiee, M.K.Akbari, A novel model for credit card fraud detection usingArtificial Immune Systems, Elsevier, Applied
Soft Computing, Volume 24, p40-49 (2014)
32. A.Mubalik (Mubarek) , E.Adali, Multilayer Perception Neural network technique for fraud detection, IEEE, Computer Science
and Engineering (UBMK), International Conference, p383-387 (2017)
33. M.Zanin, M.Romance, S.Moral, R.Criado, Credit card fraud detection through parenclitic network analysis, arXiv; 1706.01953v1, p1-8
(2017)
34. M.Tavallaee, E.Bagheri, W.Lu, A. Ghorbani, A Detailed Analysis of the KDD CUP 99Data Set, IEEE, Computational Intelligence
in Security and Defense Applications (CISDA 2009).
35. L.Dhanabal, Dr. S.P. Shantharajah, A Study on NSL-KDD Dataset for Intrusion Detection System Based on Classification
Algorithms, International Journal of Advanced Research in Computer and Communication Engineering, Vol. 4, Issue 6, p446-452 (2015)
36. Y.Dai, J.Yan, X.Tang, H.Zhao, M.Guo, Online Credit Card Fraud Detection: A Hybrid Framework with Big Data Technologies,
IEEE, Trustcom/BigDataSE/ISPA, p1644-1652 (2016)
37. O. Adewumi, A. Akinyelu, A survey of machine-learning and nature-inspired based credit card fraud detection techniques,
Springer, International Journal of System Assurance Engineering and Management, Volume 8, Supplement 2, p937–953 (2016)
38. N.Malini, M.Pushpa, Analysis on Credit Card Fraud Identification Techniques based on KNN and Outlier Detection, IEEE, Advances
in Electrical, Electronics, Information, Communication and Bio-Informatics (AEEICB), Third International Conference (2017)
39. S.Askari, A.Hussain, Credit Card Fraud Detection Using Fuzzy ID3, IEEE, Computing, Communication and Automation (ICCCA), p446-452
(2017)
40. C.Li, D.S. Poskitt, X.Zhao, The Bivariate Probit Model, Maximum Likelihood Estimation, Pseudo True Parameters and Partial Identification,
http://business.monash.edu/econometrics-and-business-statistics/research/publications, p2-34 (2016)
41. C.G. Atkeson, A.W. Moore, S. Schaal, Locally weighted learning for control, Artificial Intelligence Review 11, p1–5 (1997)