Research paper topic: A Multiple Classifiers System for Anomaly Detection in Credit Card Data With Unbalanced and Overlapped Classes Attached research paper,I need 4 ppt slides for Results and discuss

Digital Object Identifier 10.1109/ACCESS.2020.2972009 A Multiple Classifiers System for Anomaly Detection in Credit Card Data With Unbalanced and Overlapped Classes SURAYA NURAIN KALID 1 , KENG-HOONG NG 1 , GEE-KOK TONG 1 , AND KOK-CHIN KHOR 2 1 Faculty of Computing and Informatics, Multimedia University, Cyberjaya 63100, Malaysia 2 Lee Kong Chian Faculty of Engineering Science, Universiti Tunku Abdul Rahman, Kajang 43000, Malaysia Corresponding author: Kok-Chin Khor ([email protected]) This work was supported by the Ministry of Higher Education Malaysia under Grant FRGS/1/2019/SS01/MMU/03/11. ABSTRACT Frauds and default payments are two major anomalies in credit card transactions. Researchers have been vigorously nding solutions to tackle them and one of the solutions is to use data mining approaches. However, the collected credit card data can be quite a challenge for researchers. This is because of the data characteristics that contain: (i) unbalanced class distribution, and (ii) overlapping of class samples.

Both characteristics generally cause low detection rates for the anomalies that are minorities in the data.

On top of that, the weakness of general learning algorithms contributes to the difculties of classifying the anomalies as the algorithms generally bias towards the majority class samples. In this study, we used a Multiple Classiers System (MCS) on these two data sets: (i) credit card frauds (CCF), and (ii) credit card default payments (CCDP). The MCS employs a sequential decision combination strategy to produce accurate anomaly detection. Our empirical studies show that the MCS outperforms the existing research, particularly in detecting the anomalies that are minorities in these two credit card data sets. INDEX TERMS Anomaly detection, credit card, multiple classiers.

I. INTRODUCTION Credit cards are widely used because they ease our daily transactions in many ways. However, banks need to take note of these issues seriously, i.e., (i) the intervention of unauthorised third parties frauds, and (ii) the negligence of repayment by cardholders default payments. According to [1], the global credit card fraud losses have shown an uptrend, from USD 9.84 billion in the year 2011 to USD 27.69 billion in the year 2017. It is also reported that the worldwide credit card fraud is expected to reach a total of USD 31.67 million by the year 2020. The Malaysian banking sector also reported a total loss of RM 51.3 million in the credit card fraud in the year 2016 [2]. It was reported that in the year 2016, the outstanding balance of credit card holders in Malaysia is RM36.9 million and 12.8% of them failed to pay the minimum payment of the balance [3]. The Central Bank of Malaysia (Bank Negara Malaysia) also reported that The associate editor coordinating the review of this manuscript and approving it for publication was Zhe Xiao .the high outstanding balance by the credit card holders, has triggered an alarm to the Malaysian government [4]. Researchers have been vigorously nding ways to tackle both issues, including data mining. Data mining is not an option or a trend, but more of a necessity that the banking sector should invest in [5], [6]. However, banking data such as credit card fraud and default payment are quite of a challenge to data mining researchers. This is because the data usually exhibited characteristics: (i) unbalanced class distribution, and (ii) overlapping of class samples. The size of the important classes in the data, i.e., fraud and default payment, are usually the minorities. Generally, it is easy for learning algorithms to nd their regularities if they have sufcient records. But when their numbers are very small, nding their regularities becomes difcult and so as generalising their actual decision regions using learning algorithms [7] [10]. It adds difculty if their attribute values are overlapped by a large amount of normal transactions.

In general, the performance of learning algorithms will be less affected if the minority classes are linearly separable, even though the data involved are highly unbalanced [11] [13].

28210 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see http://creativecommons.org/licenses/by/4.0/ VOLUME 8, 2020 S. N. Kalid et al.: MCS for Anomaly Detection in Credit Card Data with Unbalanced and Overlapped Classes FIGURE 1.

The decision boundary that separates the majority class and minority class samples.

Another aspect to take note is the weakness of learn- ing algorithms in assessing its own classication capabil- ity [7], [11] [14]. A commonly used metric to evaluate data mining results, which computes the number of correctly clas- sied records, is the classication accuracy. The learning algorithms generally assume that positive and negative sam- ples are roughly equal in data. Therefore, many learning algo- rithms aim to maximize accuracy, which lean more towards majority classes and against minorities. Subsequently, they are unlikely to produce satisfactory results when dealing with unbalanced data sets, especially the minority classes. In a nutshell, the characteristics of the credit card data and learning algorithms have caused the problem of low anomaly detection rates. Therefore, single classiers may not give good classication results. Hence, this research aims to design a Multiple Classier System (MCS) for mitigating the low anomaly detection rate problem on the credit card data sets utilised in this study.

II. LITERATURE REVIEW A. UNBALANCED CLASS DISTRIBUTION A data set is unbalanced in its class distribution when one or more classes have a much greater number of samples than the other classes [17], [18]. In reality, the number of anomalies are very much fewer than normal transactions in the credit card data. The unbalanced class distribution problem has been the focus of many researchers [19] [22]. This is due to the high probability of producing errors in classication. The unbal- anced class distribution is as illustrated in Fig. 1. The curved line that separates between two classes is called the decision boundary, which separates the region of different classes [23]. Classifying an unbalanced data set may lead to the low classication rate problem [24] [26]. According to [27], gen- eral learning algorithms are able to give high accuracies on balanced data sets, but not on unbalanced data sets. Many researchers did comparative studies using some popular learning algorithms on unbalanced data sets [21], [28], [29].

In the next section, some selected popular single learning algorithms shall be discussed. A brief explanation on why they are weak towards unbalanced data sets shall be given. FIGURE 2.

The illustration of finding the class label for a new sample.

B. LEARNING ALGORITHM 1) NAÏVE BAYES NB is a simple, yet powerful learning algorithm that uses the probabilistic method to classify data samples. It assumes that every attribute is conditionally independent of the other attributes [30]. It will predict whether a data sample belongs to one class or another based on the Bayesian Theorem as per (1).

P.C jX /D P .X jC /P (C ) P .X / (1) Let Xbe a data sample (evidence) that is described by multiple attributes. The probability of Xbelongs to a class C is calculated as P(CjX). P(C) is the initial or prior probability while P(XjC) is the likelihood or the probability that the sample data is observed. P(X ) is the evidence with a constant value and therefore can be omitted.

According to [31], NB is weak against unbalanced data sets as it biases towards the majority class. As illustrated in Fig. 2, given two classes of different class distribution (class A 20:

class B 10) and a new sample that needs to be classied. Firstly, NB will nd the prior probability P(C) of each class; it is assumed that the new sample will belong to Class A as it has more samples as compared with Class B. Secondly, the likelihood is calculated based on the number of samples of each class that is within the vicinity (within the circle) of the new sample. Lastly, the posterior probability is calculated by combining the result of the prior probability and the likelihood of the new sample (refer to (2) and (3)).

Posterior prob :VClass A D20 30 3 20 D 0:100(2) Posterior prob :VClass B D10 30 2 10 D 0:067(3) The new sample is classied as Class A because of the largest posterior probability. Using this example, it shows that NB is likely bias towards the majority class.

2) C4.5 C4.5 is a popular learning algorithm that uses a divide- and-conquer method to build a decision tree from a training set [32] [34]. It is popular due to its ability to produce good classication results in a much shorter time. To improve classication performance, it prunes small and deep nodes VOLUME 8, 2020 28211 S. N. Kalid et al.: MCS for Anomaly Detection in Credit Card Data with Unbalanced and Overlapped Classes FIGURE 3.

The illustration of finding the class label for a new sample using KNN.

in the preliminary tree caused by the `noises' contained in training samples. The advantage of pruning is that it will decrease the risk of `over-tting' [33], [35], [36]. Over-tting refers to a classier that learns a training data, the details as well as noises, too well. Due to the inability of such classier in generalising the training data well, the classier is weak in classifying new or unknown data. Avoiding over-tting gives a more precise classication for unknown data [37], [38]. Nevertheless, the pruning pro- cess can also be a disadvantage to unbalanced data sets.

Removing `noises' from such data set may also remove small and deep nodes of the preliminary tree that belong to a minority class, thus reducing the coverage for a precise classication [14], [39] [41].

3) K-NEAREST NEIGHBOUR (KNN) KNN builds the classier's function by majority vote of its local neighbouring data points [42], [43]. Fig. 3 shows how KNN identies the class of a new sample. Suppose the number of neighbours, kD 5, and the Euclidean distance is the distance measure. The KNN classier will nd the nearest ve samples to the new sample. The Euclidean distance between the target sample ( x) and the new sample ( y) in an n-dimensional space is calculated using the measure as per (4), where pis 2.

L p D (n X i D 1 j x i y ij p )1 =p (4) Out of ve neighbours, the samples of class A is more than the samples of class B. Therefore, the new sample is classied as class A. This example shows that KNN is likely bias towards the majority class. The chance of the new sample to be classied as class B is relatively low as compared with Class A.

4) ARTIFICIAL NEURAL NETWORK (ANN) ANNs are made up of simple and highly interconnected nodes that respond to inputs by the dynamic state of the nodes [42], [44]. It is made up of layers, i.e., input layer, hid- den layer, and output layer, as shown in Fig. 4. These layers FIGURE 4.

The illustration of an ANN model that contains an input layer, hidden layers, and an output layer. FIGURE 5.

SVM transforms an original feature space into a higher dimensional feature space for finding a better decision boundary.

are formed using interconnected `nodes' that are associated with activation functions. ANNs contain some forms of learning rules that modify the weights of the connections according to the input pat- terns learning by examples [42], [45]. Similar to other single classiers, ANNs are also biased towards majority classes when they involve unbalanced data sets [26], [46].

Due to overwhelming samples of majority classes, samples of minority classes will be imperceptible to ANNs.

5) SUPPORT VECTOR MACHINE (SVM) SVM solves classication and regression problems.

As shown in Fig. 5, SVM plots each sample as a point in n -dimensional space, where nis the number of attributes in a data set. The value of each attribute will be the value of a particular coordinate. Then, SVM will identify the best hyper-plane that is able to differentiate classes [44]. SVM is powerful in getting the best decision boundary between classes. However, SVM does not work well with data sets that contain unbalanced class distribution, noises and overlapping class samples. The parameters in SVM can be altered to make the classier more immune to noises and to work well for balanced data sets. But when it involves unbalanced data sets, minority class samples may consider as noises. Therefore, minority class samples will be ignored completely by SVM [47].

C. OVERLAPPING OF CLASSES The other factor that contributes to the low classication rate is the overlapping class samples. Overlapping happens when 28212 VOLUME 8, 2020 S. N. Kalid et al.: MCS for Anomaly Detection in Credit Card Data with Unbalanced and Overlapped Classes FIGURE 6.

The unbalanced classes and overlapping samples that may lead to a low classification rate.

the samples are located too close to the decision boundary of classes and overlapped with each other [20], [21]. Based on a systemic study using a set of articially gen- erated data sets prepared by [48], the study result showed that the degree of overlapping class samples had a strong correlation with unbalanced class distribution. Fig. 6 shows the relationship between unbalanced class distribution and overlapping class samples. When the samples of two classes are not well distributed, some samples may overlap with each other. The grey area in Fig. 6 indicates the samples that are overlapped. Another reason of overlapping is because the samples in both classes share almost that same value of attributes. Such overlapping causes difculties for a classier to classify the samples and may eventually lead to a low classication rate [49], [50]. In the next section, solutions to the unbalanced class distribution and the overlapping class samples shall be discussed based on the previous work conducted by other researchers.

D. MULTIPLE CLASSIFIERS SYSTEM (MCS) Many researchers attempted different solutions to address the low classication rate problem caused by the two problems, as discussed in the earlier section [51] [54], [83], [84]. One of the solutions is MCS. MCS is the combination of a set of classiers to produce a better prediction. MCS employs various decision combination strategies that are able to pro- duce a more robust, reliable, efcient recognition and accu- rate classication [55] [59]. There are three combination strategies when employing MCS: (i) sequential combination, (ii) parallel combination, and (iii) hybrid combination.

1) SEQUENTIAL COMBINATION Using the sequential combination, two or more single classi- ers process the input data in a sequential manner. As shown in Fig. 7, the output resulted from a single classier will then be used as the input data for the subsequent single classiers. Usually, simple classiers are utilised rst, followed by more accurate and complex classiers [55]. However, this order can be reversed depending on the needs of the design.

When the prior classier is unable to accurately classify one of the class samples, then the sample is given to the next classier for further classication [58]. An example FIGURE 7.

The sequential combination of MCS. The MCS passes data from one classifier to another.

of learning algorithm that uses sequential combination is Boosting [60]. Reference [61] used AdaBoost on a credit card fraud data set. The data set is highly unbalanced with only 0.173% fraud transaction. The authors used Naïve Bayes as the base classier. Upon completing the experiment, they obtained 0.999, 0.825 and 0.981 for accuracy, true positive rate and true negative rate, respectively. However, accuracy is not a good performance measure as compared to the other two measures when involving unbalanced data sets. This is because the accuracy measure favours the majority class.

Reference [62] used a ne-tuned boosting ensemble approach, which is known as XGBoost, to solve a classica- tion problem. The problem was to decide on the granting of loan application. The data sets used in the experiments were German credit, Australian credit, Taiwan credit, P2P landing data set A, and P2P landing data set B. Three of them are unbalanced and the other two are approximately balanced in class distribution. The classier's performance was measured using accuracy and the highest accuracy obtained was 0.879.

Using accuracy in their work showed that the authors did not explicitly address the issues of unbalanced data and overlap- ping class samples. However, the authors suggested to inte- grate XGBoost into MCS to further improve the classication performance. MCS is one of the recommended approaches to consider in data mining research when involving unbalanced data sets. Reference [79] also used boosting algorithm, AdaBoost.

M1. The authors boosted three different classiers, which were Multilayer Perceptron (MLP), Radial Basis Func- tion (RBF), and Naïve Bayes (NB), on the Taiwan credit card default payment data. To handle the unbalanced data set issue, the authors reduced the size of majority class samples by using random under sampling.

2) PARALLEL COMBINATION The same data are processed by multiple single classi- ers using the parallel combination, where each classier is VOLUME 8, 2020 28213 S. N. Kalid et al.: MCS for Anomaly Detection in Credit Card Data with Unbalanced and Overlapped Classes FIGURE 8.

The parallel combination of the MCS. All the single classifiers involved are independent from each other.

independent from the others. The output from all the single classiers will be combined to get a nal decision, as shown in Fig. 8. An example of learning algorithm that uses parallel combination is Bagging [60]. Reference [63] used Bagging with Random Forest as the base classier. The data set used was about credit card default payment from a Taiwan bank. The data set is unbalanced with 28% of it categorised as default-payment. The author obtained 0.816 for accuracy, 0.371 for TPR, and 0.764 for AUC. Reference [64] used the bagged decision trees to classify new transactions as fraudulent or legitimate type. The data set used in the experiment was the UCSD-FICO competition credit card data set. Out of 97,707 instances, the data set contains only 2.9% of fraudulent instances. The authors did some comparison with other single classiers and the result showed that bagging approach outperformed other single classiers.

The authors of [78] used Random Forest, an ensemble method. Random Forest uses the same combination strategy as Bagging. Similar to [63], credit card default payment data set from a Taiwan bank was used. In the experiment, the authors applied the Correlation Based Feature Selec- tion (CFS) technique to reduce the data set dimension with the purpose of improving classication accuracy. Based on the experiment result, the authors obtained a good TPR f 0.816. In the study by [81], the credit card fraud data set used was obtained from a European bank with an unbalanced class dis- tribution and overlapping class samples. The data set is highly unbalanced with only 492 fraud transactions (0.173%). The authors rst partitioned and clustered the data set. Then, a Random Forest framework, with C4.5 as the base classier, was trained using the resulted clusters. The nal result was determined based on the majority vote. The authors used AUC as the performance measure. Based on the experiment result, the author obtained an AUC value of 0.965. The study by [82] used the similar data set used as [81].

The data set was bootstrapped such that each resulted data set had a balanced class distribution. Then, the authors used an ensemble of Deep Belief Network and applied to each of the bootstrapped samples. The author used the performance measures, i.e., accuracy, TPR, TNR, and AUC and obtained 0.906, 0.818, 0.995, and 0.978, respectively. FIGURE 9.

The MCS that utilises the hybrid combination. The sequential and parallel combinations are put into one architecture.

3) HYBRID COMBINATION The hybrid combination puts the sequential and parallel com- binations into one architecture. Fig. 9 illustrates the hybrid combination. In this case, the input data is fed to the rst classier. The output from the rst classier will be the input to several parallel classiers. Then, a single combination function or classier will merge the output of the individual parallel classiers.

Reference [65] used the hybrid combination to classify credit or loan applicants into good and bad applicants. The data set used in the experiment is a German credit data set. The data set has 20 attributes with 700 Good and 300 Bad Applicant data. The hybrid combinations used are a two-level voting scheme: level I AdaBoost approach, and level II single classier approach. The authors used accuracy as the performance measure. Both level I and level II had achieved an average accuracy of 76.33% and 78.33%, respectively. However, the evaluation measure is not suitable as the study involved the unbalanced data set.

Reference [80] also employed a hybrid model using a combination of AdaBoost and majority voting. ANN and NB were used as the base classiers for AdaBoost that employed sequential combination. Then, the nal result was obtained using majority voting, which was done in parallel. Similar to [81], the credit card fraud data set obtained from the European bank was used in this experiment. Upon completing the experiment, the authors were able to obtain an accuracy of 0.999, a TPR of 0.789 and a TNR 0.999.

E. EVALUATION MEASURES Given a binary class problem, general learning algorithms assume that the classes involved are approximately bal- anced and attempt to maximise its accuracy regardless of classes [66], [67]. With a balanced data set, accuracy is the suitable measure to evaluate the performance of a classier.

However, when it involves an unbalanced data set, learning algorithms may bias towards the majority class [68] [70].

This may lead to a high accuracy in overall, but a poor classication rate for the minority class.

In this study, the True Positive Rate (TPR) was used to evaluate the performance of the proposed MCS. By using 28214 VOLUME 8, 2020 S. N. Kalid et al.: MCS for Anomaly Detection in Credit Card Data with Unbalanced and Overlapped Classes TABLE 1.

The attributes of the Credit card fraud (CCF) data set. TPR as the evaluation metric, we were able to identify the classication rate for both majority and minority classes.

In this study, we focused on the minority classes: (i) the credit card frauds and (ii) the credit card default payments.

III. METHODOLOGY A. DATA SETS OVERVIEW AND THEIR INHERITED PROBLEMS Two credit card data sets were utilised in this study to demonstrate the challenges: (i) overlapping class samples and (ii) unbalanced class distribution. The general learning algorithms have difculty in handling these two issues and caused low detection rates for minority classes. The rst data set is the Credit card fraud (CCF) data set released by [71]. Credit card fraud is an act of gaining unlaw- ful advantage such as performing a variety of unauthorised transactions using the victim's credit card account [72], [73].

This data set was formed during a big data mining and fraud detection research between Worldline and the Machine Learning Group of Université Libre de Bruxelles (ULB).

The data set contains transactions made by European credit card holders. The data set has a total of 284,807 transac- tions and it is highly unbalanced with only 492 fraud trans- actions (0.173%). Further, the data set has a total number of 31 attributes, as shown in Table 1. Unfortunately, due to condentiality, the details of certain attribute are not dis- closed. Most attributes of the CCF data set exhibit the sce- narios as follows. Fig. 10 (a) shows the pairwise relationship of attributes V20 and V14. The red crosses are samples of Class 1 (frauds), while the blue crosses represent samples of Class 0 (normal transactions). The class distribution is clearly unbalanced and the samples of both classes are overlapped.

Fig. 10 (b) also displays the same scenario where the plot involves attributes V24 and V12.

The second data set is the credit card default pay- ment (CCDP) data set [74], [75]. Default credit card payment refers to the failure of a credit card holder in performing the minimum amount of credit card repayment within the agreed period [76], [77]. This data set contains the payment data of credit card holders of a Taiwan bank from April 2005 to September 2005. The CCDP data set is also slightly unbalanced with a ratio of approximately 3:1 (non-default payment: default payment). It has a total of 30,000 payment instances. 23,364 instances belong to `no' class (non-default payment next month), and 6,636 of them belong to the `yes' class (default payment next month). This means that there are only 28% of default payment instances out of the whole FIGURE 10.

The visualisation of the CCF data set showing the unbalanced class distribution and the overlapping class problem between attributes (a) V20 vs. V14, and (b) V24 vs. V12.

payment data. This data set has a total of 25 attributes. The detail of each attribute is described in Table 2.

Fig. 11 (a) shows the scatter plot of attributes ``Repay- ment status in April'' and ``Amount of previous state- ment in April''. The red crosses are samples of class `yes' (non-default payment), while the blue crosses are samples of class `no' (default payment). The distribution of both classes is clearly unbalanced and we can see that the blue crosses are overwhelmed by the red crosses. Fig. 11 (b) shows the scatter plot of attributes ``Bill statement in August'' and ``Repayment status in August''. Fig. 11 (a) and (b) show the overlapping samples of major- ity and minority classes. It is expected to be difcult for clas- siers to accurately detect the minority classes (class 1 and class `yes').

B. TACKLING THE PROBLEM USING MCS MCS is the combination of predictions from a set of classiers to produce a better prediction. MCS employs decision combination strategies that are able to produce a more robust, reliable, efcient recognition and accurate classication [55] [58], [85]. Basically, there are three different combination strate- gies when employing MCS: (i) sequential, (ii) parallel, and (iii) hybrid. Among the three combination strategies of MCS, VOLUME 8, 2020 28215 S. N. Kalid et al.: MCS for Anomaly Detection in Credit Card Data with Unbalanced and Overlapped Classes TABLE 2.

The attributes of the CCDP data set. we expect the sequential combination to perform well. Using the sequential combination, the output of the rst classier will be an input of the subsequent classier. This means that the same piece of sample will be classied more than once.

Taking credit card fraud data set (CCFD) as the example, our primary concern will be the frauds that are misclassi- ed as normal transactions. Therefore, the sample classied as normal by the rst classier will be fed to the subse- quent classier for re-classication purpose. Further, having one classier for each class is an advantage. Since each of the classes shall be handled by one classier, then it will mitigate the effect of the unbalanced class distribution and FIGURE 11.

The visualisation of the CCDP data set showing the unbalanced class distribution and overlapping class problem between attributes (a) ``Amount of previous statement in April'' vs. ``Repayment status in April'', and (b) ``Repayment status in August'' vs. ``Amount of bill statement in August''.

overlapping classes. Employing the MCS with sequential combination will therefore reduce the overall misclassi- cation rate and improve the detection rate, particularly for the minority class. Algorithm 1 shows the pseudocode of the proposed MCS that utilises the sequential combination scheme.

The credit card data set shall be classied by classier C1 (Line 1, Algo. 1). Classier C1is the expert classier to classify majority class samples. During the classication, if the sample is classied as 1 or `yes', then the sample will be stored in the data set F_ds. Conversely, if the sample is classied as 0 or `no', then the sample will be stored in a different data set called N_ds(Line 2 8, Algo. 1). Our pri- mary concern for this project is the samples of class 1 or `yes' that are misclassied as 0 or `no'. To address the concern, we will re-classify the data by feeding N_dsinto classier C2 (Line 9, Algo. 1). Classier C2is the expert classier to classify minority class samples. During the classication, if the sample is classied as 1 or `yes', then the sample will be stored in data set F_ds. If the sample is classied as 0 or `no', then the sample will remain in the data set N_ds (Line 10 14, Algo. 1). Subsequently, we combine F_dsand N_ds into one data set called C_ds(Line 15, Algo. 1).Finally, a confusion matrix as per Table 3 shall be generated based on C_ds (Line 16 19, Algo. 1). Once the confusion matrix 28216 VOLUME 8, 2020 S. N. Kalid et al.: MCS for Anomaly Detection in Credit Card Data with Unbalanced and Overlapped Classes Algorithm 1 Detect_Anomaly (CC_db ) Input:

The credit card data set, CC_dbthat comprises of xfeatures for each credit card transaction Output: A confusion matrix of classied data 1.classify CC_dbwith classier C1//C1 classier to classify // majority class samples 2.for each transaction, sin the CC_db do 3.if C1. class( s) is equal to 1 or `yes' // it is a fraud transaction 4.assign sto F_ds //F_ds is a data set to keep 1 or `yes' data 5.else { C1.Class( s) is or `no'} // it is a normal transaction 6.assign sto N_ds //N_ds is a data set to keep or `no' data 7.end if 8.end for 9.classify N_dswith classier, C2//C2 classier to classify minority // class samples 10.for each transaction, snin the N_dsdo 11.if C2. class( sn) is equal to 1 or `yes' 12.assign sntoF_ds 13.end if 14.end for 15.combine F_dsandN_ds assign to C_ds//C_ds are combination of // dataset F_ds and N_ds 16. actual DC_ds //actual class 17. predicted DC_ds //predicted class 18. result confusion matrix ( actual,predicted) 19.return result 20. TPRmin Calculate TPRmin( result)// calculate the TPR for // minority class samples 21. TPRmaj Calculate TPRmaj( result)// calculate the TPR for // majority class samples 22.return TPRmin, TPRmaj TABLE 3.

The Confusion matrix for the classified data. is obtained, the True Positive Rate (TPR) for both majority and minority classes shall be calculated using (5) and (6) (Line 20 22, Algo. 1). To implement Algorithm 1, we need to identify the expert classiers, C1and C2.An experiment had been conducted to identify them. A few popular single classiers, namely, Naïve Bayes (NB), C4.5, Random Forest, Random Tree, Logistic Regression (LR), Multilayer Perceptron (MLP), and IBk were tested on both CCF and CCDP data sets. The experiments were evaluated using TPR. The TPR for both majority and minority classes can be calculated using (5) and (6).

TPR .minority /D Num :of detected frauds or default payments Total frauds or default payments (5) TPR .majority /D Num :of detected non frauds or non default payments Total non frauds or non default payments (6) Apart from TPR, Area under the ROC Curve (AUC) were also used as the performance metric. In general, AUC tells the TABLE 4.

The experimental results of using single classifiers. on both CCF and CCDP data sets. capability of a classier in differentiating classes. The closer the AUC value to 1, the better a classier is in distinguishing between classes.

As shown in Table 4, most classiers including C4.5 were able to achieve the perfect TPR for the majority class (class 0) of the CCF data set. C4.5 also scored the highest TPR, 0.955, for the majority class (class `no') of the CCDP data set. Therefore, we identied C4.5 as the C1as it is able to produce high TPRs for the majority class of both data sets.

On the other hand, NB obtained the highest TPR, 0.829, for the minority class (class 1) of the CCF data set. NB also obtained the highest TPR for the minority class (class yes), 0.719, of the CCDP data set. Therefore, we identied NB as C2 as it is able to produce high TPR for the minority class of both data sets. We employed sequential combination in our proposed MCS, with C4.5 and NB as the rst and second expert clas- siers. The results shall be discussed in the next section.

IV. RESULTS & DISCUSSION A. DETECTION RESULTS USING SINGLE CLASSIFIERS Based on our studies in the literature review, it was found that single classiers are weak against classifying data sets that contain unbalanced class distribution and overlapping classes. This is proven by our experiment using some popular single classiers, as shown in Table 4.

The majority class of the CCF data set was classied per- fectly, with a TPR of 1.000, by most of the single classiers.

As for the minority class, the TPRs were mostly just average.

The highest TPR for class 1 is 0.829, which was achieved by using NB. On the other hand, the TPRs for the majority class of the CCDP data set were quite promising, except for NB which VOLUME 8, 2020 28217 S. N. Kalid et al.: MCS for Anomaly Detection in Credit Card Data with Unbalanced and Overlapped Classes TABLE 5.

The comparison between the proposed MCS and the other researchers' work on the CCF data set. TABLE 6.

The comparison between the proposed MCS and the other researchers' work on the CCDP data set. achieved only 0.634. As for the minority class, the TPRs were generally low, except for NB that scored an average TPR of 0.719. In general, the single classiers did not perform well in detecting the minority class in both CCF and CCDP data sets.

B. DETECTION RESULTS USING THE PROPOSED MCS Table 5 shows the comparison between our approach and the other researcher' approach in classifying CCF data set.

All the other researchers' work listed above used ensem- ble approaches in tackling the unbalanced class distribution.

By using our proposed MCS, we managed to achieve the highest TPR of 0.872 for the minority class and outperformed the other researchers'' work. Our proposed approach also gave a good accuracy of 0.999 and a TNR of 1.000. Our MCS was also tested on CCDP data set. Table 6 shows the comparison between our approach and the other researchers' work. We outperformed their work by obtain- ing the highest TPR for the minority class, which is 0.840.

Our proposed approach also achieved an accuracy of 0.930 and a TNR of 0.955, which are better than their work. In summary, we can conclude that our proposed approach is able to tackle the unbalanced class distribution and the overlapping class samples that exists in both credit card data sets.

V. CONCLUSION Credit card is one alternative of cash payment. Some card holders may abuse their responsibility in credit card usage and repayment. Apart from that, credit card transaction is also prone to fraudulent where unauthorized parties perform illegal transactions using credit cards. Therefore, it is the responsibility of card issuers or the banks to nd an effective way to reduce the cost that may incur when the issues above happen. One way to address these issues is via data mining.

Due to the characteristics such as overlapping class samples and unbalanced class distribution that exist in credit card data sets, it gives challenges to data mining researchers. On top of that, the weakness of general learning algorithms also contributes to the difculties of classifying the minority class, which is usually the important class, of the data sets. This study proposed a MCS to tackle the issues as dis- cussed above. Based on our analysis using single classiers, we found that C4.5 is the expert in classifying the majority class samples and NB is the expert in classifying the minor- ity class samples. Therefore, they were arranged sequentially in our proposed MCS to detect credit card anomalies. Our proposed MCS was evaluated using two different credit card data sets: CCF and CCDP. We have compared our work with the other researchers' work. The experimental results showed that the proposed MCS outperformed their work.

In general, our proposed MCS demonstrates its superiority in handling the credit data sets that inherit the characteris- tics of overlapping classes and unbalanced class distribution.

However, there are rooms to improve the TPR for the minor- ity classes. We are looking into other MCS combination strategies for our future work, particularly the hybrid com- bination. Currently, researchers had attempted deep learn- ing algorithms such as Long Short-term Memory (LSTM) and Deep Belief Networks for detecting anomalies in credit card transactions [86]. We are also considering combining the deep learning algorithms, as in the study of [87], for promising detection results.

REFERENCES [1] Nilsonreport . Accessed: 2019. [Online]. Available: https://nilsonreport.

com/upload/content_promo/The_Nilson_Report_10-17-2016.pdf [2] RM51.3mil in payment card fraud losses reported in 2016 . Accessed: 2019.

[Online]. Available: https://www.thestar.com.my/news/nation/2017/08/03/ rm422mil-in-credit-card-fraud-losses-reported-in-2016 28218 VOLUME 8, 2020 S. N. Kalid et al.: MCS for Anomaly Detection in Credit Card Data with Unbalanced and Overlapped Classes [3]T. Lokman.

3.6 Million Credit Card Holders Have RM36.9 Billion Out- standing Balance . NST Online. Accessed: Dec. 2019. [Online]. Available:

https://www.nst.com.my/news/nation/2017/08/270620/36-million-credit- card-holders-have-rm369-billion-outstanding-balance [4]Ceicdata. (2019). Malaysia Credit Card Statistics . Accessed: Dec. 2019.

[Online]. Available: https://www.ceicdata.com/en/malaysia/credit-card- statistics [5]A. O. Adewumi and A. A. Akinyelu, ``A survey of machine-learning and nature-inspired based credit card fraud detection techniques,'' Int.

J. Syst. Assur. Eng. Manag. , vol. 8, no. S2, pp. 937 953, Nov. 2017, doi:10.1007/s13198-016-0551-y.

[6]S. Makki, ``Fraud analysis approaches in the age of big data A review of state of the art,'' in Proc. IEEE 2nd Int. Workshops Found. Appl. Self Syst.

(FASW) , 2017, pp. 243 250, doi:10.1109/fas-w.2017.154.

[7]G. M. Weiss, ``Mining with rarity,'' SIGKDD Explor. Newsl., vol. 6, no. 1, p. 7, Jun. 2004, doi:10.1145/1007730.1007734.

[8]S. Maheshwari, J. Agrawal, and S. Sharma, ``A new approach for classi- cation of highly imbalanced datasets using evolutionary algorithms,'' Int.

J. Sci. Eng. Res. , vol. 2, no. 7, pp. 1 5, 2011.

[9]Q. Dong, S. Gong, and X. Zhu, ``Class rectication hard min- ing for imbalanced deep learning,'' in Proc. IEEE Int. Conf. Com- put. Vis. (ICCV) , Oct. 2017, pp. 1851 1860. Accessed: Aug. 2018, doi:10.1109/iccv.2017.205.

[10]J. Gao, L. Gong, J. Y. Wang, and Z. C. Mo, ``Study on unbalanced binary classication with unknown misclassication costs,'' in Proc. IEEE Int. Conf. Ind. Eng. Eng. Manage. (IEEM) , Dec. 2018, pp. 1538 1542, doi:10.1109/ieem.2018.8607671.

[11]N. Japkowicz and S. Stephen, ``The class imbalance problem: A systematic study1,'' IDA, vol. 6, no. 5, pp. 429 449, Nov. 2002, doi:10.3233/ida- 2002-6504.

[12]J. Mathew, C. K. Pang, M. Luo, and W. H. Leong, ``Classication of imbal- anced data by oversampling in kernel space of support vector machines,'' IEEE Trans. Neural Netw. Learn. Syst. , vol. 29, no. 9, pp. 4065 4076, Sep. 2018, doi:10.1109/tnnls.2017.2751612.

[13]P. Vuttipittayamongkol, E. Elyan, A. Petrovski, and C. Jayne, ``Overlap- based undersampling for improving imbalanced data classication,'' in Proc. Intell. Data Eng. Automated Learn. (IDEAL) , 2018, pp. 689 697, doi:10.1007/978-3-030-03493-1_72.

[14]N. Chawla, N. Japkowicz, and A. Kotcz, ``Editorial,'' ACM SIGKDD Explor. Newslett. , vol. 6, no. 1, p. 1, 2004, doi:10.1145/1007730.1007733.

[15]A. Fernández, S. Del Río, N. V. Chawla, and F. Herrera, ``An insight into imbalanced Big Data classication: Outcomes and challenges,'' Complex Intell. Syst. , vol. 3, no. 2, pp. 105 120, Jun. 2017, doi:10.1007/s40747- 017-0037-9.

[16]V. H. Barella, L. P. F. Garcia, M. P. De Souto, A. C. Lorena, and A. De Carvalho, ``Data complexity measures for imbalanced classication tasks,'' in Proc. Int. Joint Conf. Neural Netw. (IJCNN) , Jul. 2018, pp. 1 8, doi:10.1109/ijcnn.2018.8489661.

[17]G. Haixiang, L. Yijing, J. Shang, G. Mingyun, H. Yuanyue, and G. Bing, ``Learning from class-imbalanced data: Review of methods and applications,'' Expert Syst. Appl., vol. 73, pp. 220 239, May 2017, doi:10.1016/j.eswa.2016.12.035.

[18]A. G. De Sá, A. C. Pereira, and G. L. Pappa, ``A customized classication algorithm for credit card fraud detection,'' Eng. Appl. Artif. Intell., vol. 72, pp. 21 29, Jun. 2018, doi:10.1016/j.engappai.2018.03.011.

[19]Y. Yong, ``The research of imbalanced data set of sample sampling method based on K-means cluster and genetic algorithm,'' Energy Procedia, vol. 17, pp. 164 170, 2012, doi:10.1016/j.egypro.2012.02.078.

[20]J. Bªaszczy«ski and J. Stefanowski, ``Neighbourhood sampling in bagging for imbalanced data,'' Neurocomputing, vol. 150, pp. 529 542, Feb. 2015, doi:10.1016/j.neucom.2014.07.064.

[21]K. Napieraªa, J. Stefanowski, and S. Wilk, ``Learning from imbalanced data in presence of noisy and borderline examples,'' in Rough Sets and Current Trends in Computing . Brookline, MA, USA: Microtome Publishing, 2010, pp. 158 167, doi:10.1007/978-3-642-13529-3_18.

[22]G. Lemaitre, F. Nogueira, and C. Aridas, ``Imbalanced-learn: A Python toolbox to tackle the curse of imbalanced datasets in machine learning,'' J. Mach. Learn. Res. , vol. 18, no. 1, pp. 1 5, 2017.

[23]A. Engelbrecht and H. Viktor, Rule Improvement Through Deci- sion Boundary Detection Using Sensitivity Analysis (Lecture Notes in Computer Science). Berlin, Germany: Springer, 1999, pp. 78 84, doi:10.1007/bfb0100474. [24]Y. Qi.

A Brief Literature Review of Class Imbalanced Prob-LEM .

Accessed: Mar. 2019. [Online]. Available: http://nyc.lti.cs.cmu.edu/IRLab/ 11-743f04/qyj/IR-pages/ImbalancedSummary.html [25]M. Goldstein and S. Uchida, ``A comparative evaluation of unsupervised anomaly detection algorithms for multivariate data,'' PLoS ONE, vol. 11, no. 4, Apr. 2016, Art. no. e0152173, doi:10.1371/journal.pone.0152173.

[26]B. Krawczyk, ``Learning from imbalanced data: Open challenges and future directions,'' Prog. Artif. Intell., vol. 5, no. 4, pp. 221 232, Nov. 2016, doi:10.1007/s13748-016-0094-0.

[27]S. Boughorbel, F. Jarray, and M. El-Anbari, ``Optimal classier for imbalanced data using Matthews correlation coefcient metric,'' PLoS ONE , vol. 12, no. 6, Jun. 2017, Art. no. e0177678, doi:10.1371/journal.

pone.0177678.

[28]I. Brown and C. Mues, ``An experimental comparison of classication algorithms for imbalanced credit scoring data sets,'' Expert Syst. Appl., vol. 39, no. 3, pp. 3446 3453, Feb. 2012, doi:10.1016/j.eswa.2011.09.033.

[29]T. Fitzpatrick and C. Mues, ``An empirical comparison of classication algorithms for mortgage default prediction: Evidence from a distressed mortgage market,'' Eur. J. Oper. Res., vol. 249, no. 2, pp. 427 439, Mar. 2016, doi:10.1016/j.ejor.2015.09.014.

[30]T. Li, J. Li, Z. Liu, P. Li, and C. Jia, ``Differentially private naive Bayes learning over multiple data sources,'' Inf. Sci., vol. 444, pp. 89 104, May 2018, doi:10.1016/j.ins.2018.02.056.

[31]G. Nanda, K. Vallmuur, and M. Lehto, ``Improving autocoding perfor- mance of rare categories in injury classication: Is more training data or ltering the solution?'' Accident Anal. Prevention , vol. 110, pp. 115 127, Jan. 2018, doi:10.1016/j.aap.2017.10.020.

[32]S. Salzberg, ``C4.5: Programs for machine learning by J. Ross Quinlan.

Morgan Kaufmann Publishers, Inc., 1993,'' Mach. Learn., vol. 16, no. 3, pp. 235 240, 1994, doi:10.1023/a:1022645310020.

[33]S. Sharma, J. Agrawal, and S. Sharma, ``Classication through machine learning technique: C4. 5 algorithm based on various entropies,'' Int. J.

Comput. Appl. , vol. 82, no. 16, pp. 28 32, Nov. 2013, doi:10.5120/14249- 2444.

[34]H. Sani, C. Lei, and D. Neagu, Computational Complexity Analysis of Decision Tree Algorithms (Lecture Notes in Computer Science). Cham, Switzerland: Springer, 2018, pp. 191 197, doi:10.1007/978-3-030-04191- 5_17.

[35]J. Crowley. (2017). Intelligent Systems: Reasoning and Recogni- tion . Accessed: Mar. 2018. [Online]. Available: http://www-prima.imag.

fr/Prima/jlc/Courses/2016/ENSI2.SIRR/ENSI2.SIRR.S1.pdf [36]S. Selvi, S. Sowmiya, and R. Sangeetha. (2018). C5 Causal Deci- sion Tree . Accessed: Mar. 2018. [Online]. Available: http://ijsrcseit.com/ paper/CSEIT1833100.pdf [37]J. Lever, M. Krzywinski, and N. Altman, ``Model selection and over- tting,'' Nature Methods , vol. 13, no. 9, pp. 703 704, Sep. 2016, doi:10.1038/nmeth.3968.

[38]R. A. Watson and E. Szathmáry, ``How can evolution learn?'' Trends Ecol. Evol. , vol. 31, no. 2, pp. 147 157, Feb. 2016, doi:10.1016/j.tree.2015.11.009.

[39]W. Liu, S. Chawla, D. A. Cieslak, and N. V. Chawla, ``A robust decision tree algorithm for imbalanced data sets,'' in Proc. SIAM Int. Conf. Data Mining , Apr. 2010, pp. 766 777, doi:10.1137/1.9781611972801.67.

[40]A. Dal Pozzolo and G. Bontempi, ``Adaptive machine learning for credit card fraud detection,'' Ph.D. dissertation, Université Libre de Bruxelles, Bruxelles, Belgium, 2015.

[41]F. Westerlund and P. Wijayatunga, ``Credit card fraud detectioN (machine learning algorithms),'' M.S. thesis, Ume Univ., Krong Battambang, Cambodia, 2017.

[42]O. Maimon and L. Rokach, Data Mining and Knowledge Discovery Hand- book . Boston, MA, USA: Springer, 2010.

[43]R. Taghizadeh-Mehrjardi, K. Nabiollahi, B. Minasny, and J. Triantalis, ``Comparing data mining classiers to predict spatial distribution of USDA-family soil groups in Baneh region, Iran,'' Geoderma, vols. 253 254, pp. 67 77, Sep. 2015, doi:10.1016/j.geoderma.

2015.04.008.

[44]D. Tien Bui, T. A. Tuan, H. Klempe, B. Pradhan, and I. Revhaug, ``Spatial prediction models for shallow landslide hazards: A comparative assess- ment of the efcacy of support vector machines, articial neural networks, kernel logistic regression, and logistic model tree,'' Landslides, vol. 13, no. 2, pp. 361 378, Apr. 2016, doi:10.1007/s10346-015-0557-6.

[45]R. Nazir, E. Momeni, K. Marsono, and H. Maizir, ``An articial neural network approach for prediction of bearing capacity of spread foundations in sand,'' J. Teknol., vol. 72, no. 3, p. 1, 2015, doi:10.11113/jt.v72.

4004.

VOLUME 8, 2020 28219 S. N. Kalid et al.: MCS for Anomaly Detection in Credit Card Data with Unbalanced and Overlapped Classes [46]Y. L. Murphey, H. Guo, and L. A. Feldkamp, ``Neural learning from unbalanced data,'' Int. J. Speech Technol. , vol. 21, no. 2, pp. 117 128, Sep. 2004, doi:10.1023/b:apin.0000033632.42843.17.

[47]R. Akbani, S. Kwek, and N. Japkowicz, ``Applying support vector machines to imbalanced datasets,'' Mach. Learn., vol. 2004, pp. 39 50, Mar. 2004, doi:10.1007/978-3-540-30115-8_7.

[48]G. Batista, R. Prati, and M. Monard, Balancing Strategies and Class Over- lapping (Lecture Notes in Computer Science). Berlin, Germany: Springer, 2005, pp. 24 35, doi:10.1007/11552253_3.

[49]S. Ullman and E. Sali, ``Object classication using a fragment-based repre- sentation,'' in Biologically Motivated Computer Vision . Berlin, Germany:

Springer, 2000, pp. 73 87, doi:10.1007/3-540-45482-9_8.

[50]Z. Yang and D. Gao, ``Classication for imbalanced and overlapping classes using outlier detection and sampling techniques,'' Appl. Math.

Inf. Sci. , vol. 7, no. 1L, pp. 375 381, Feb. 2013, doi:10.12785/amis/ 071l50.

[51]Y. Sahin and E. Duman, ``Detecting credit card fraud by decision trees and support vector machines,'' in Proc. Int. MultiConf. Eng. Comput. Scientists , Hong Kong, 2011, pp. 1 6.

[52]K. R. Seeja and M. Zareapoor, ``FraudMiner: A novel credit card fraud detection model based on frequent itemset mining,'' Sci. World J., vol. 2014, pp. 1 10, May 2014, doi:10.1155/2014/252797.

[53]M. Fahmi, A. Hamdy, and K. Nagati, ``Data mining techniques for credit card fraud detection: Empirical study,'' in Proc. Int. Conf. Sustain. Vital Technol. Eng. Inform. , 2016, pp. 1 9.

[54]A. Pawar, S. Dongare, A. Deokate, H. Sangle, and P. Mokal, ``Outlier detection using oversampling PCA for credit card fraud detection,'' Int.

J. Res. Appl. Sci. Eng. Technol. , vol. 5, no. 5, pp. 1 11, 2017.

[55]A. F. R. Rahman and M. C. Fairhurst, ``Multiple classier decision com- bination strategies for character recognition: A review,'' Int. J. Document Anal. Recognit. , vol. 5, no. 4, pp. 166 194, Jul. 2003, doi:10.1007/s10032- 002-0090-8.

[56]J. Stefanowski. (2008). Lamsade.Dauphine.fr. Accessed: Aug. 2018.

[Online]. Available: http://www.lamsade.dauphine.fr/~projet_cost/ ALGORITHMIC_DECISION_THEORY/pdf/Stefanowski/Stefanowski4.

pdf [57]M. Wo¹niak, M. Graña, and E. Corchado, ``A survey of multiple classier systems as hybrid systems,'' Inf. Fusion, vol. 16, pp. 3 17, Mar. 2014, doi:10.1016/j.inffus.2013.04.006.

[58]A. A. Aburomman and M. B. I. Reaz, ``A novel SVM-kNN-PSO ensem- ble method for intrusion detection system,'' Appl. Soft Comput., vol. 38, pp. 360 372, Jan. 2016, doi:10.1016/j.asoc.2015.10.011.

[59]H. Sayadi, N. Patel, S. P. D. , A. Sasan, S. Rafatirad, and H. Homayoun, ``Ensemble learning for effective run-time hardware-based malware detection: A comprehensive analysis and classication,'' in Proc. 55th ACM/ESDA/IEEE Design Automat. Conf. (DAC) , 2018, Art. no. 1, doi:10.1109/dac.2018.8465828.

[60]J. Brownlee. (2019). Bagging and random forest ensemble algorithms for machine learning. Machine Learning Mastery. Accessed: Dec. 2018.

[Online]. Available: https://machinelearningmastery.com/bagging-and- random-forest-ensemble-algorithms-for-machine-learning/ [61]I. Sohony, R. Pratap, and U. Nambiar, ``Ensemble learning for credit card fraud detection,'' in Proc. ACM India Joint Int. Conf. Data Sci. Manage.

Data (CoDS-COMAD) , 2018, pp. 289 294. Accessed: Dec. 21, 2019, doi:10.1145/3152494.3156815.

[62]Y. Xia, C. Liu, Y. Li, and N. Liu, ``A boosted decision tree approach using Bayesian hyper-parameter optimization for credit scoring,'' Expert Syst.

Appl. , vol. 78, pp. 225 241, Jul. 2017, doi:10.1016/j.eswa.2017.02.017.

[63]P. Singh, ``Comparative study of individual and ensemble methods of clas- sication for credit scoring,'' in Proc. Int. Conf. Inventive Comput. Inform.

(ICICI) , Nov. 2017, pp. 968 972, doi:10.1109/icici.2017.8365282.

[64]M. Zareapoor and P. Shamsolmoali, ``Application of credit card fraud detection: Based on bagging ensemble classier,'' Procedia Comput. Sci., vol. 48, pp. 679 685, Dec. 2015, doi:10.1016/j.procs.2015.04.201.

[65]M. Abedini, F. Ahmadzadeh, and R. Noorossana, ``Customer credit scor- ing using a hybrid data mining approach,'' Kybernetes, vol. 45, no. 10, pp. 1576 1588, Nov. 2016, doi:10.1108/k-09-2015-0228.

[66]A. Ali, S. Shamsuddin, and A. Ralescu. (2015). Classication With Class Imbalance Problem: A Review . Accessed: Aug. 2018.

[Online]. Available: http://home.ijasca.com/data/documents/13IJASCA- 070301_Pg176-204_Classication-with-class-imbalance-problem_A- Review.pdf [67]S. Wang, L. L. Minku, and X. Yao, ``A systematic study of online class imbalance learning with concept drift,'' IEEE Trans. Neural Netw. Learn. Syst. , vol. 29, no. 10, pp. 4802 4821, Oct. 2018, doi:10.1109/tnnls.2017.2771290.

[68]M. Bekkar, H. Djemaa, and T. Alitouche, ``Evaluation measures for models assessment over imbalanced data sets,'' J. Inf. Eng. Appl., vol. 3, no. 10, Apr. 2013.

[69]H. Hossin and M. N. Sulaiman, ``A review on evaluation metrics for data classication evaluations,'' Int. J. Data Mining Knowl. Manage. Process , vol. 5, no. 2, pp. 1 11, Mar. 2015, doi:10.5121/ijdkp.2015.5201.

[70]J. Akosa, ``Predictive accuracy: A misleading performance measure for highly imbalanced data,'' in Proc. SAS Global Forum, 2017, pp. 2 5.

[71]A. D. Pozzolo, O. Caelen, R. A. Johnson, and G. Bontempi, ``Cali- brating probability with undersampling for unbalanced classication,'' in Proc. IEEE Symp. Ser. Comput. Intell. , Dec. 2015, pp. 159 166, doi:10.1109/ssci.2015.33.

[72]S. Maes, K. Tuyls and B. Vanschoenwinkel, ``Credit card fraud detection using Bayesian and neural networks,'' in Proc. 1st Int. Naiso Congr. Neuro Fuzzy Technol. , 2002, pp. 261 270.

[73]N. Carneiro, G. Figueira, and M. Costa, ``A data mining based system for credit-card fraud detection in e-tail,'' Decis. Support Syst., vol. 95, pp. 91 101, Mar. 2017, doi:10.1016/j.dss.2017.01.002.

[74]K. Bache and M. Lichman. (2013). UCI Machine Learning Reposito-ry . Accessed: Mar. 2018. [Online]. Available: http:// archive.ics.uci.edu/ml/index.php [75]I. Yeh. (2016). UCI Machine Learning Repository: Default of Credit Card Clients Data Set . Accessed: Mar. 2018. [Online]. Available:

https://archive.ics.uci.edu/ml/datasets/default+of+credit+card+clients [76]L. Irby. (2018). Learn How Credit Card Default Happens and What You Can Do About It . Accessed: Mar. 2018. [Online]. Available:

https://www.thebalance.com/what-is-credit-card-default-960209 [77] Kuala Lumpur , Credit Card, Bank Negara Malaysia, Kuala Lumpur, Malaysia, 2019.

[78]A. Venkatesh, and S. Gracia, ``Prediction of credit-card defaulters: A comparative study on performance of classiers,'' Int. J. Comput. Appl., vol. 145, no. 7, pp. 36 41, 2016, doi:10.5120/ijca2016910702.

[79]A. Charleonnan, ``Credit card fraud detection using RUS and MRN algo- rithms,'' in Proc. Manage. Innov. Technol. Int. Conf. (MITicon) , Oct. 2016, pp. 1 73, doi:10.1109/miticon.2016.8025244.

[80]K. Randhawa, C. K. Loo, M. Seera, C. P. Lim, and A. K. Nandi, ``Credit card fraud detection using AdaBoost and majority voting,'' IEEE Access, vol. 6, pp. 14277 14284, 2018, doi:10.1109/access.2018.2806420.

[81]H. Wang, P. Zhu, X. Zou, and S. Qin, ``An ensemble learning framework for credit card fraud detection based on training set partitioning and clustering,'' in Proc. IEEE SmartWorld, Ubiquitous Intell. Comput., Adv. Trusted Comput., Scalable Comput. Commun., Cloud Big Data Comput., Internet People Smart City Innov. (Smart- World/SCALCOM/UIC/ATC/CBDCom/IOP/SCI) , 2018, pp. 94 98, doi:10.1109/smartworld.2018.00051.

[82]P. Xenopoulos, ``Introducing DeepBalance: Random deep belief net- work ensembles to address class imbalance,'' in Proc. IEEE Int. Conf.

Big Data (Big Data) , Dec. 2017, pp. 3684 3689, doi:10.1109/big- data.2017.8258364.

[83]A. Salazar, G. Safont, A. Soriano, and L. Vergara, ``Automatic credit card fraud detection based on non-linear signal processing,'' in Proc. IEEE Int. Carnahan Conf. Secur. Technol. (ICCST) , Oct. 2012, pp. 207 212, doi:10.1109/ccst.2012.6393560.

[84]A. Salazar, G. Safont, and L. Vergara, ``Semi-supervised learning for imbalanced classication of credit card transaction,'' in Proc. Int. Joint Conf. Neural Netw. (IJCNN) , Jul. 2018, pp. 1 7, doi:10.1109/ijcnn.2018.8489755.

[85]A. Soriano, L. Vergara, B. Ahmed, and A. Salazar, ``Fusion of scores in a detection context based on alpha integration,'' Neural Comput., vol. 27, no. 9, pp. 1983 2010, Sep. 2015, doi:10.1162/neco_a_00766.

[86]A. Roy, J. Sun, R. Mahoney, L. Alonzi, S. Adams, and P. Beling, ``Deep learning detecting fraud in credit card transactions,'' in Proc. Syst. Inf. Eng.

Design Symp. (SIEDS) , 2018, pp. 129 134.

[87]D. Bermejo-Peláez, S. Ash, G. Washko, R. S. J. Estépar, and M. Ledesma- Carbayo, ``Classication of interstitial lung abnormality patterns with an ensemble of deep convolutional neural networks,'' Sci. Rep., vol. 10, no. 1, pp. 1 5, 2020, doi:10.1038/s41598-019-56989-5.

28220 VOLUME 8, 2020 S. N. Kalid et al.: MCS for Anomaly Detection in Credit Card Data with Unbalanced and Overlapped Classes SURAYA NURAIN KALID was born in Kuala Terengganu, Terengganu, Malaysia, in 1990.

She received the Diploma degree in IT and the bachelor's degree in IT majoring in software engi- neering from Multimedia University, Malaysia, in 2010 and 2014, respectively, where she is cur- rently pursuing the master's degree. From 2014 to 2016, she was an Assistant Manager Core Network, responsible for assisting the upgrading of IP infrastructure for Telekom Malaysia's backbone. Then, she was an IT Executive assigned to TM's subsidiary Fiberail under the Network Management Centre. Since November 2016, she has been an Assistant Lecturer with Multimedia Uni- versity, Malaysia. Her research interests include articial intelligence, data mining, and software development. KENG-HOONG NG is currently a Lecturer with the Faculty of Computing and Informat- ics, Multimedia University. He has been with the University, as an Academic Staff, since 2002.

He enjoys lecturing, provides training (CCNA professional course), and doing research. For his personal interest, he likes to travel, exercise (badminton), and nd undervalued stocks in his spare time. His research interests include data clustering, data classication, bioinformatics, and computer networks. GEE-KOK TONG currently holds a lecturer posi- tion with the Faculty of Computing and Informat- ics, Multimedia University, Cyberjaya, Malaysia.

He has been conducting research work in nancial econometric time series, such as vector error cor- rection model, GARCH modeling, extreme value theory, copula, value-at-risk, conditional value-at- risk, risk adjusted performance measures, portfolio management, and Monte Carlo simulation. KOK-CHIN KHOR received the B.Sc. and M.Sc.

degrees from Universiti Putra Malaysia (UPM), and the Ph.D. degree from Multimedia Univer- sity (MMU), Malaysia. Apart from teaching, he is also a certied Instructor for Cisco professional networking courses. His research interests include data mining and computer networking. Prior to joining the academic sector, he worked as a Web Programmer and System Analyst. He was the Chair of a few special sessions of international research conferences, as well as the reviewer for a numbers of journals and conferences. VOLUME 8, 2020 28221