See discussions, stats, and author profiles for this publication at: [Link]

net/publication/359635597

Developing a Credit Card Fraud Detection Model using Machine Learning

Approaches

Article in International Journal of Advanced Computer Science and Applications · April 2022
DOI: 10.14569/IJACSA.2022.0130350

CITATIONS READS
22 4,490

5 authors, including:

Abdullah Alourani Ashraf Ali

Qassim University Arab Open University - Bahrain
40 PUBLICATIONS 394 CITATIONS 30 PUBLICATIONS 236 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Ashraf Ali on 31 March 2022.

The user has requested enhancement of the downloaded file.

 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 13, No. 3, 2022

Developing a Credit Card Fraud Detection Model

using Machine Learning Approaches
Shahnawaz Khan1 Bharavi Mishra3
Faculty of Engineering Department of Computer Science & Engineering
Design and Information & Communications Technology The LNM Institute of Information Technology
Bahrain Polytechnic Jaipur, Rajasthan, India
Isa Town, Bahrain

Ashraf Ali4
Abdullah Alourani2 Faculty of Computer Studies
Department of Computer Science and Information Arab Open University, Kingdom of Bahrain
College of Science in Zulfi
Majmaah University
Al-Majmaah, 11952 Mustafa Kamal5
Saudi Arabia College of Science & Theoretical Studies
Saudi Electronic University, Dammam, Saudi Arabia

Abstract—The growing application and usage of e-commerce further categorized into offline fraud, application fraud,
applications have given an exponential rise to the number of bankruptcy fraud, internal fraud, behavioural fraud, counterfeit
online transactions. Though there are several methods for fraud, cardholder-not-present fraud, etc. Online transactions are
completing online transactions, however, credit cards are most providing new opportunities for fraudsters. Frauds are activities
commonly used. The increased number of transactions has given by the fraudsters that are intended to yield the fraudster
the opportunity to the fraudsters to mislead the customers and personal or financial gain [23]. These activities are often
make them execute fraudulent transactions. Therefore, there is a criminal or wrongful. Credit card frauds cause problems and
need for such a method that can automatically classify detect losses to financial institutions as well as individuals. There are
fraudulent transactions. This research study aims to develop a
hundreds of transactions every second for any financial
credit-card fraud detection model that can effectively classify an
online transaction as fraudulent or genuine. Three supervised
institution [2]. Manual fraud detection and prevention is not a
machine learning approaches have been applied to develop a feasible solution. There has been a tremendous amount of
credit-card fraud classifier. These techniques include logistic effort by the research community in developing efficient
regression, artificial intelligence and support vector machine. detection techniques for credit card frauds. So that the trust of
The classification accuracy achieved by all the classifiers is the customers can increase in e-commerce and online
almost similar. This research has used the confusion matrix and transactions and the losses that occurred due to the frauds can
area under the curve to demonstrate the score of the different be minimized.
performance measures and evaluate the overall performance of
Digital transactions can take place over the phone or on the
the classifiers. Several performance measures such as accuracy,
precision, recall, F1-measure, Matthews correlation coefficient,
internet. For executing a transaction, very basic information is
receiver operating characteristic curve have been computed and required such as expiry date, card number, card verification
analysed to evaluate the performance of the credit-card fraud number etc. Cardholders provide this information through
detection classifiers. The analysis demonstrates that the support phone or the internet. Fraudsters apply several techniques and
vector machine-based classifier outperforms the other classifiers. attempt to steal the credit card information of the customers so
that they can use it for doing fraudulent transactions. It is a
Keywords—Credit card fraud detection; neural network; very serious, and costly problem for financial service
support vector machine; logistic regression; performance measures providers. Billions of dollars are subject to fraudulent
transactions every year [22]. The fraudulent transaction is an
I. INTRODUCTION issue of concern for all the credit card providers or by
With the increased use of financial technology, the use of expansion for all the financial systems that provide the
online transactions has increased manifolds in recent years. facilities for online transactions to their customers. It is usually
This expansion and use of electronic commerce have increased the result of someone stealing the credit card information of the
the trust of customers in online transactions. There are several customers which also impact the brand value of the credit card
kinds of financial fraud such as credit card fraud, securities service providers and the merchants.
fraud, insurance fraud, etc. that use online methods to The worldwide cost of fraudulent transactions is projected
accomplish the fraud. Most online transactions utilize credit to be 38.5 billion U.S. dollars by 2027 and it was 32 billion
cards. Therefore, the most common type of fraud among all the dollars in 2021 [26]. The fraudulent transactions cause a huge
frauds types is credit card fraud [17]. Credit card frauds can be

411 | P a g e
[Link]
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 13, No. 3, 2022

loss for the merchants also because they usually have to bear classification models. Among the implemented models, the
all the related costs such as administrative charges, issuer fees support vector machine model performs better.
etc. The number of digital transactions is huge, therefore,
verifying each transaction for its genuineness is not an easy This paper has been organized into five sections. The
task for the financial service providers. Consequently, credit second section of the paper examines the literature review and
card providers often only investigate the cases when they are presents the background work briefly. Section three discusses
reported by the customers. The literature highlights several machine learning techniques, data and pre-processing. Section
other issues in which the primary issue is the imbalance four presents the results obtained from different classifiers
amount of the cases in the available historical data. The implemented by this research study and also exhibits the
number of actual fraud cases in the data is usually very small in various performance measures and evaluates the performance
comparison to the number of genuine transactions. The of the credit card fraud detection models. The last section
imbalance of the training data creates the problem of biases in presents the conclusion.
the classification accuracy of the classifier. The presence of the II. RELATED WORK
dominating class corners the other classes. Thus, the classifier
keeps predicting the dominating class. Therefore, even if the In binary classification, the basic concept is to find the
classifier is predicting wrong, the accuracy of the classifier will threshold value that enables the classifier to assign a particular
not be impacted by large. label to the case or instance being predicted. There have been
conducted several research studies on detecting credit card
Therefore, there is a need for a system that can detect frauds based on the spending behaviour of the customer.
fraudulent transactions efficiently and raise an alarm as soon as However, sometimes the customer spending behaviour changes
the transaction is made, so that the credit card provider can take during certain conditions such as holidays or other special
immediate action and reduce the risk of capital loss. Several occasions. This might create an issue for supervised machine
researchers have utilized multiple machine learning and other learning systems. Research by [3] proposes a hybrid approach
computational methods to detect credit card frauds. This for managing customer abnormal spending behaviour. The
research aims to tackle the issue of data imbalance and develop proposed approach combines supervised and unsupervised
a credit card fraud detection system. There are several machine learning approaches and presents effective results in
techniques that aim to minimize the effect of the data case of abnormalities of the spending behaviours.
imbalance. This research applies such techniques for data
preparation and during the model evaluation phase. Because A research study [7] applies a generative adversarial
the researchers are in the view that only measuring the network (GAN) to detect credit card frauds. As the credit card
accuracy will not be a proper evaluation of the classifier. transaction data is usually highly imbalanced, the proposed
framework by the research study [7] has a higher false-positive
Various researchers have applied several machine learning rate if the sensitivity is improved. Research [22] applies and
and hybrid methods for detecting fraudulent transactions and compares several machine learning techniques such as Naïve
have developed classifiers that can detect fraudulent Bayes, random forest, logistic regression, decision tree,
transactions. Several researchers have used standalone methods AdaBoost, multiple layer perceptron etc. The research
[7], [8], [13], [17], [27] while many researchers have also demonstrates that the AdaBoost with majority voting produces
applied hybrid approaches [3], [22] for detecting credit card the best results among all the alternatives. There is an
frauds. The issue of the presence of the highly imbalanced interesting fact to consider for this research [22], that the
amount of data samples is the primary challenge in developing transactions data used for developing the classifier has only
an effective credit-card fraud detection model. Several 0.0355% fraudulent transactions. The data used is highly
approaches have been applied such as feature selection, feature imbalanced and no measures have been taken to counter the
engineering, sequence classification, supervised and imbalances of the data. In the research study presented by [17],
unsupervised machine learning methods, data pre-processing to several machine learning methods have been applied for credit
balance the data classes. card fraud detection. The study illustrates that SVM, ANN,
Credit card fraud detection methods discussed in this C5.0 decision tree, and LR performs better among the tested
research focus on identifying if a transaction on a credit card is criteria. However, the number of false positives is high among
fraudulent or not by applying several machine learning all the implemented methods.
techniques such as logistic regression, artificial neural A random forest algorithm is an effective method for
networks, and support vector machines. Credit card fraud developing supervised classification models. A research study
detection systems use historical transaction data to train. The by [27] implements random forest supervised machine learning
decision of these systems relies on the spending behavioural techniques to detect the behavioural patterns for genuine
patterns learned during the training process from historical transactions and fraudulent transactions. A similar research
transaction data. The system aims to develop an efficient study by [13] proposes a random forest algorithm-based
credit-card fraud detection model that can effectively classify machine learning model for detecting credit card frauds. The
the transactions into genuine transactions or fraudulent model presented by the research [13] exhibits good accuracy.
transactions efficiently. Several performance measures such as Though, the performance measure is based on the statistics
accuracy, precision, recall, F1-measure, Matthews correlation obtained from the confusion matrix only. As has been
coefficient, receiver operating characteristic curve have been discussed above by several researchers the credit card fraud
calculated to evaluate the performance of different detection problem poses the challenge of the imbalance data
classes. Therefore, there should have been some other

412 | P a g e
[Link]
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 13, No. 3, 2022

performance measure such as the receiver operating III. METHODOLOGY

characteristic (ROC) curve that could have been employed for
a better performance measure. In a similar domain, the research A. Dataset and Pre-processing
study by [18] focuses on minimizing the number of incorrect One of the primary issues for data preparation for credit-
fraud classifications. Actually, that’s the primary target of any card fraud transaction data is the labelling of the data. Often the
researcher working in this domain. The research study by [18] fraudulent label of the transactional data can only be decided
employs multiple algorithms for anomaly detection and posterior the transaction has been executed and reported by the
implements algorithms such as isolation forest and local outlier customer. The dataset used in this research study consists of
factor algorithms. The results presented by the study are transactions made by European cardholders in September 2013
sensitive to the quantity of the data and face the challenge of [14]. The dataset contains 284,807 transactions made during
imbalanced data for the classification classes. two days. The fraudulent transactions made during this time
were 492 which is just 0.172% of all transactions made during
A research study by [20] presents an interesting perspective this time. As it is evident that the data is imbalanced, therefore,
on credit card fraud detection and infer that there is no constant this research uses the resampling technique and makes an effort
pattern for fraud. Therefore, supervised machine learning to oversample the fraudulent transactions and to remove the
techniques are not efficient in credit card fraud detection. It genuine transactions. The dataset was transformed using
[20] proposes an unsupervised machine learning approach principal component analysis to maintain the confidentiality of
using a restricted Boltzmann machine (RBM) and deep Auto- the transactions [14] and the principal components are used as
encoder. However, the results achieved are less promising than features for training the classifiers. The dataset contains 30
some of the supervised machine learning approaches. A input features such as transaction time, transaction amount and
Bayesian network classifier-based approach presented by [6] in 28 principal components. The output classes have two labels 1
a research study uses a hyper-heuristic evolutionary algorithm and 0. The fraudulent transactions are assigned label 1 and the
to detect the patterns. The presented solution like the other genuine transactions are labelled as 0.
approaches discussed in this section targets the class imbalance
and misclassification issues of the credit card fraud detection B. Modelling for Credit Card Fraud Detection
problem. Research by [2] introduces a real-time fraud detection Several approaches and algorithms have been implemented
system using machine learning and big data. This solution for credit card fraud detection. Some of these solutions have
primarily focuses on the detection speed of the transactions, been discussed in the related work section. Several features and
use of big data and scalability. affairs have been taken into consideration for credit card fraud
The research study by [4] considers the spatial and detection classifiers. One of the common issues that are
temporal features among others and presents a 3D discussed throughout the literature in credit card fraud
convolutional neural network for credit card fraud detection. detection is the presence of class data imbalance. However,
The present approach [2] implements the model on the real- none of the algorithms or approaches discussed precisely tackle
world data collected from multiple locations. The research by the class imbalance issue. Therefore, this research implements
[15] proposes a hidden Markov model-based approach for a two-step process for handling the issue. The first phase is
automated feature engineering to improve the performance of data pre-processing. In the data pre-processing phase, the study
the classifier and to model temporal feature correlation. A aims to reduce the class data imbalance by increasing the
similar research study [28] that focus on the features of the number of cases for the minority class, and by reducing the
transactional data, develops a deep learning-based solution that number of cases for dominating class. This section discusses
uses homogeneity-oriented behaviour analysis for feature the approaches used in this study.
engineering. A research study [1] proposes an optimized light 1) Logistic regression: Logistic regression is a
gradient boosting based machine learning technique for probabilistic modelling process that produces the probability
predicting credit card frauds. This research [1] relies on
of the discrete output variables based on the input variables.
parameter optimization for improving the performance of the
classifier. Often logistic regression is applied for binary classification
when the input variable is single or multiple. However,
An interesting approach to solving the credit card detection logistic regression can be applied to classify more than two
problem is sequence classification or prediction problem. output classes, which is known as multinomial logistic
Research [8] formulates credit card fraud detection as a
regression. Furthermore, it can be used for ordering the level
sequence classification problem. It applies long short-term
memory neural network to identify the fraudulent transaction. of the output variable classes which is known as ordered
The research concludes that articulating the fraud detection logistic regression. However, logistic regression is often used
task as a sequence-learning problem leads to an increased for binary classification problems. Credit card fraud detection
number of false positives. As a matter of fact, online is a binary classification problem in which the output of the
transactions should not be considered a sequential transaction is either fraud or a genuine transaction given the
classification problem, because the amount, time, and point of input features for the transactions. Therefore, logistic
the online transactions usually change randomly. It will require regression can be used as a credit card fraud detection
a highly disciplined spending behaviour to express online technique.
transactions as a sequence learning problem.

413 | P a g e
[Link]
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 13, No. 3, 2022

Credit card detection can be performed by computing the architectures that are employed in machine learning for
probability of the given transaction using the given features various different tasks. The following diagram (Fig. 1)
and comparing it with a threshold value such as 0.5. If the presents a general feed-forward neural network architecture.
computed probability is more than 0.5 then it will be classified This study employs a feed-forward neural network. Input layer
as fraud if less than the threshold then it will be classified as a
is the first layer and works as input to the neural network. The
genuine transaction. Let us assume that the probability of the
fraudulent transaction based on the transaction features x is P(y input layer of the neural network used in this study contains
= 1|x) or simply P(x). To compute the probability-estimate log- 25 neurons corresponding to the features of a transaction. Two
odds can be computed. Log-odds are directly proportional to middle layers, popularly known as hidden layers, have been
the probability of the transaction label. Higher the odds, the used each of which contains 10 neurons. The output layer has
higher the probability of the given label for the transaction. 2 neurons corresponding to each class. The network uses the
backpropagation algorithm for learning. The layers are fully
It can be defined as: . (1) connected layers. The activation function used is rectified
linear units (ReLU).
For modelling and simplifying the computational process,
natural logarithm was applied as follows: Input layer Hidden layers Output layer
( ) (2)

Let’s consider, ( ) (3)

Here, is the transpose of the weight vector and b is the

offset variable. The above equation (Eq. 3) can be further
simplified by applying exponential on both sides:

...…
...…
(4)

Therefore, the probability of the fraudulent can be

estimated using the above equation (Eq. 4) in which x are the
features of the transaction. The aim is to optimize the values of
w and b based on the transactional data. It can be learnt by Fig. 1. Feed-forward Neural Network General Architecture.
converting the above problem into maximum likelihood
estimation problem and optimizing it for w and b using the The ReLU activation function is a very simple but effective
transactional data. activation function. It returns a zero if the input received by the
activation function is negative, otherwise, no change is applied
The log-likelihood from the equation (eq 4) can be derived on the input and the input values is returned as it is if the input
as following (Eq 5): is positive. It can be simply stated as:
∑ (5)
{ (6)
The optimized values of w and b are estimated by
maximizing the log-likelihood (Eq. 5) or by converting the 3) Support vector machines: Support vector machines like
above problem into minimization problem after multiplying
artificial neural networks have been among the most popular
with a negative sign.
machine learning algorithms. SVMs are commonly applied in
2) Artificial neural network: Artificial neural network solving supervised machine learning problems such as
(ANN) is one of the most powerful machine-learning regression, classification and outlier detection. Though, the
techniques. ANN aims to simulate the behaviour of biological number of samples for this research study is ample, but, SVMs
organisms. In the human nervous system, neurons are can also be applied in a scenario where the number of
connected to other neurons through connections which are dimensions is more than the sample size. SVMs, for
known as axons and dendrites. The strength of the connections classification, functions by finding the best hyperplanes that
is subject to change in accordance with the external stimuli separate the data points in accordance with the classes. The
which are referred to weights in ANN. The computational hyperplanes set apart the data points of one class from the data
units in ANN are termed as neurons. Though, the ANN points of other classes. The optimization is applied for finding
simulation of the biological organism is very basic still, the the best hyperplane that can find the maximum margin among
complexity and computation capabilities of the artificial the data points of one class from the data points of the other
neural network is very powerful. Artificial neural networks class. The support vectors are the closes data points to the
have been applied to solve complex computational problems hyperplane. SVMs can be implemented using different types
for example in machine translation [9], [10], [11], [24], [25], of kernel functions. Kernel methods are a set of algorithms
image processing [12], time series forecasting [21], that are used in machine learning techniques for pattern
classification etc. There are several neural network analysis and detection. Kernel functions, transforms the data

414 | P a g e
[Link]
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 13, No. 3, 2022

into higher dimensions, expecting to find clearer decision The models (logistic regression, artificial neural network
boundaries for data separation. Kernel functions aids in and support vector machines) achieved an accuracy of 99.91%,
efficiently transforming high dimensional data for creating 99.91% and 99.94% respectively.
optimal boundaries for decision making. The kernel function 2) Precision: Precision is also known as the positive
used in this research is the quadratic kernel function. predictive value and is the ratio of the true positive predictions
A quadratic kernel is a non-stationary and special form of over the total positive predictions:
the polynomial kernel. The general form of the polynomial
function looks as follows:
(7) The models (logistic regression, artificial neural network
Here, f and g are the computed vectors of features from the and support vector machines) achieved a precision of 87.32%,
input data samples, c is a free parameter and has a value of c ≥ 76.91% and 87.67% respectively.
0, and d is the degree of polynomial. When the degree d = 2. 3) Recall: Recall measures the true positive rate of the
Then, the kernel function is called quadratic kernel function
classifier and is also known as sensitivity in binary
and can be presented as follows:
classification. It is calculated as the fraction of the true
(8) positive predictions over all the positive cases that were
retrieved for the testing:
IV. RESULTS AND DISCUSSION
This research study implements several machine learning
methods for credit card fraud detection. The selected methods
are logistic regression, artificial neural network and support The sensitivity or recall measured for logistic regression
vector machines. Three models were trained using the above- classifier is 61.59%, recall for the artificial neural network is
mentioned machine learning technique on the selected training 75.81% and for support vector machines is 78.05%.
data. The training process has applied five-cross validation.
Test data samples were randomly selected before applying TABLE I. CONFUSION MATRIX AND PERFORMANCE METRICS FOR
resampling techniques to evaluate the performance of the LOGISTIC REGRESSION
system on real-world data. The accuracy achieved for the
logistic regression method is 99.92% and the prediction speed Logistic Regression
is around 300 thousand predictions per second. The accuracy Predicted Class
for the neural network-based model is 99.92% while the Genuine Fraud
prediction speed is 650 thousand predictions per second.
Support vector machines model has the prediction speed of Genuine 284271 44 99.98%
True Class
about 350 thousand predictions per second and the accuracy of Fraud 189 303 61.59%
99.94%. The prediction speed of the artificial neural network- 87.32% 99.92%
based model is the fastest among the tested models, but the
application of the model on big data has yet to be tested [2]. As
TABLE II. CONFUSION MATRIX AND PERFORMANCE METRICS FOR ANN
the accuracy achieved by all the developed models is similar,
therefore, some other evaluation metrics must be used to Artificial Neural Network
measure the performance of the developed models. The
Predicted Class
confusion matrix for classification models demonstrates the
true positive (TP), true negatives (TN), false positives (FP), Genuine Fraud
and false negatives (FN). It illustrates how many of the Genuine 284203 112 99.96%
instances have been classified to their actual class and how True Class
Fraud 119 373 75.81%
many have been misclassified. The following Tables I to III
illustrate the confusion matrix and several performance metrics 76.91% 99.92%
for the three classification models:
TABLE III. CONFUSION MATRIX AND PERFORMANCE METRICS FOR SVM
1) Accuracy: Accuracy is the primary performance
evaluation metric and measures the ratio of correct prediction Support Vector Machine
over the total number of predictions by the classifier. It can be Predicted Class
presented as:
Genuine Fraud
Genuine 284261 54 99.98%
True Class
Fraud 108 384 78.05%
87.67% 99.94%

415 | P a g e
[Link]
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 13, No. 3, 2022

4) Specificity: To measure the true negative prediction evaluating the performance of the classification model. This
rate of the classifier, specificity is calculated as the proportion research has considered the Receiver Operating Characteristic
of the true negative predictions over the total negative cases (ROC) curve as an additional performance measure for the
that were retrieved for the testing: classifiers. ROC curve was initially developed and applied
during the world war II for detecting the enemy objects [16].
ROC is fundamentally a graphical representation or a plot that
illustrates the accuracy of the classification capability of a
The negative prediction rate or the specificity of the logistic binary classifier (Lusted, 1971). ROC curve is a widely used
regression is 99.98%, specificity for the artificial neural performance measure to evaluate the performance of the
network is 99.96%, and the specificity for the support vector binary classifiers. ROC curve plots the sensitivity of the
machines-based classifier was measured as 99.98%. classifier against the false positive rate. False positive rate can
5) F1-Score: F1-score or F-measure considers the be obtained by subtracting the specificity of the classification
importance of true positive and true negative. It is the model from one. The graph is drawn on a 1x1 space which
harmonic mean of the two performance measures calculated means that the scale on each of the x and y-axis is in the range
earlier which are precision and recall: of 0 to 1. The line connecting the coordinates (0, 0) and (1, 1)
will represent a random classifier. An ideal classifier would
score a point on the upper left corner (0, 1) which represents
the case of zero false positives and zero false negatives.
F1-score measured for logistic regression classifier is The following Fig. 2 demonstrates the ROC curve for
72.23%, 76.36% for artificial neural network classifier, and logistic regression model. The logistic regression classification
82.58% for support vector machines. model yields a point (0.38, 1) and the area under curve is 0.97.
6) Matthews Correlation Coefficient (MCC): Matthews The following Fig. 3 demonstrates the ROC curve for
correlation coefficient [19] calculates the correlation between artificial neural network classifier. The artificial neural network
the actual classes and predicted classes of the cases. Matthews model yields the threshold point (0.24, 1) and the area under
correlation coefficient (MCC) provides a more accurate the curve for artificial neural network model is 0.90. The
evaluation of the overall performance of the binary classifier following Fig. 4 illustrates the ROC curve for support vector
than other performance measures such as precision, recall, F1- machine model. The support vector machine classification
score, and accuracy [5]. MCC is the ration of the covariance model yields a point (0.22, 1) and the area under curve is 0.94.
of the actual classes of the cases and the predicted labels over It can be seen that the point yield by the support vector
the product of the standard deviations of the true classes ( ) machine vector model is the closest to the point of the best
classification model.
and the predicted classes ( . MCC is measured as following:

√
MCC has been calculated for all the three classifiers. The
value of the Matthews correlation coefficient remains between
-1 and +1. Higher the value of MCC, better the model is. The
MCC value for logistic regression is 0.733 or 73.3%, 76.32%
for the artificial neural network classifier and 82.69% for the
support vector machine classifier.
7) Receiver operating characteristic curve: Another
important criterion to consider is the highly imbalanced
amount of data points in the training data. Highly imbalanced
data introduces several issues in developing machine learning
models. One of such issues is biasness. In the case of highly
imbalanced data, the prediction accuracy is usually biased. In
the case of imbalanced data, the accuracy calculated based on
the confusion matrix might be misleading because it will not
address the issue of biased classification. Therefore, some
other evaluation measure should also be considered while Fig. 2. ROC Curve for Logistic Regression Model.

416 | P a g e
[Link]
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 13, No. 3, 2022

supervised machine learning approaches have produced better

results. There are several issues while developing the credit-
card fraud detection model. Availability of the highly
imbalanced class data is the issue of major concern. The
presence of the dominating class corners the other classes.
Thus, the classifier keeps predicting the dominating class.
Therefore, even if the classifier is predicting wrong, the
accuracy of the classifier will not be impacted by large. This
research study has applied a resampling technique to counter
the effect of imbalanced class data. However, due to the nature
of the problem, it is neither feasible nor practical to completely
ignore and eliminate the gap of imbalanced data classes. This
research study has implemented three machine learning
techniques which are logistic regression (LR), artificial neural
network (ANN), and support vector machines (SVM).
The models have been evaluated thoroughly using different
performance evaluation measures and matrices. Though based
on the accuracy computed from the confusion matrix, all the
model scores same. But, further analysis using different
performance measures demonstrates that the support vector
machines classification model outperforms the other models.
The prediction accuracy and specificity are almost the same for
Fig. 3. ROC Curve for Artificial Neural Network Model. all the classification models, while the precision is almost 12%
lower for the ANN model than the other two models. While the
SVM model has slightly higher precision than the LR model.
Recall of the SVM model is almost 21% higher than the LR
model and almost 3% higher than the ANN model. Similarly,
the MCC value and F1-score for the SVM model are over 12%
higher than the LR model and 7% higher than the ANN model.
Receiver operating curve yields a point (0.38, 1) for the LR
model, (0.24, 1) for the ANN model, and (0.22, 1) for the SVM
model. The best-case scenario for the classifier on the ROC
curve is to yield a point on the upper left corner (0, 1) which
represents the case of zero false positives and zero false
negatives. Among, the three tested models, SVM is the closest
classification model to the best point (0, 1). Therefore, it can be
concluded based on the various performance measures that the
SVM model outperforms the other models for credit card
detection.
ACKNOWLEDGMENT
The authors would like to thank the Deanship of Scientific
Research at Majmaah University for supporting this work
under project no: R-2022-50.
REFERENCES
[1] Altyeb, A. A., & Malebary, S. J. (2020). An Intelligent Approach to
Credit Card Fraud Detection Using an Optimized Light Gradient
Fig. 4. ROC Curve for Support Vector Machine Model. Boosting Machine. IEEE Access, 1–1. doi:10.1109/access.2020.
2971354.
V. CONCLUSION [2] Carcillo, F., Dal Pozzolo, A., Le Borgne, Y.-A., Caelen, O., Mazzer, Y.,
& Bontempi, G. (2018). SCARFF: A scalable framework for streaming
Credit card fraud is an issue of concern among financial credit card fraud detection with spark. Information Fusion, 41, 182–194.
institutions and causes huge financial losses for service doi:10.1016/[Link].2017.09.005.
providers. Fraudulent transactions have cost over 32 billion [3] Carcillo, F., Le Borgne, Y.-A., Caelen, O., Kessaci, Y., Oblé, F., &
United States dollars worldwide in 2021. This amount is Bontempi, G. (2019). Combining Unsupervised and Supervised
Learning in Credit Card Fraud Detection. Information Sciences.
projected to increase by over 38 billion dollars in the next 5 doi:10.1016/[Link].2019.05.042.
years by 2027. Several computational approaches have been [4] Cheng, D., Xiang, S., Shang, C., Zhang, Y., Yang, F., & Zhang, L.
employed to develop an effective model for credit card fraud (2020). Spatio-Temporal Attention-Based Neural Network for Credit
detection. Researchers have employed supervised and Card Fraud Detection. Proceedings of the AAAI Conference on
unsupervised machine learning approaches. However, the Artificial Intelligence, 34(01), 362–369. doi:10.1609/aaai.v34i01.5371.

417 | P a g e
[Link]
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 13, No. 3, 2022

[5] Chicco, D., & Jurman, G. (2020). The advantages of the Matthews [17] Makki, S., Assaghir, Z., Taher, Y., Haque, R., Hacid, M.-S., &
correlation coefficient (MCC) over F1 score and accuracy in binary Zeineddine, H. (2019). An Experimental Study with Imbalanced
classification evaluation. BMC genomics, 21(1), 1-13. Classification Approaches for Credit Card Fraud Detection. IEEE
[6] De Sá, A. G., Pereira, A. C., & Pappa, G. L. (2018). A customized Access, 1–1. doi:10.1109/access.2019.2927266.
classification algorithm for credit card fraud detection. Engineering [18] Maniraj, S. P., Saini, A., Ahmed, S., & Sarkar, S. (2019). Credit card
Applications of Artificial Intelligence, 72, 21-29. fraud detection using machine learning and data science. International
[7] Fiore, U., De Santis, A., Perla, F., Zanetti, P., & Palmieri, F. (2017). Journal of Engineering Research and, 8(09).
Using generative adversarial networks for improving classification [19] Matthews, B. W. (1975). Comparison of the predicted and observed
effectiveness in credit card fraud detection. Information Sciences. secondary structure of T4 phage lysozyme. Biochimica et Biophysica
doi:10.1016/[Link].2017.12.030. Acta (BBA) - Protein Structure, 405(2), 442–451. [Link]
[8] Jurgovsky, J., Granitzer, M., Ziegler, K., Calabretto, S., Portier, P.-E., [Link]/10.1016/0005-2795(75)90109-9.
He-Guelton, L., & Caelen, O. (2018). Sequence classification for credit- [20] Pumsirirat, A., & Yan, L. (2018). Credit card fraud detection using deep
card fraud detection. Expert Systems with Applications, 100, 234–245. learning based on auto-encoder and restricted boltzmann machine.
doi:10.1016/[Link].2018.01.037. International Journal of advanced computer science and applications,
[9] Khan, S. N., & Usman, I. (2019). Amodel for english to urdu and hindi 9(1), 18-25.
machine translation system using translation rules and artificial neural [21] Qader, M. R., Khan, S., Kamal, M., Usman, M., & Haseeb, M. (2021).
network. Int. Arab J. Inf. Technol., 16(1), 125-131. Forecasting CO2 Emissions Due To Electricity Power Generation In
[10] Khan, S., & Mishra, R. B. (2011). Translation rules and ANN based Bahrain.
model for English to Urdu machine translation. INFOCOMP Journal of [22] Randhawa, K., Loo, C. K., Seera, M., Lim, C. P., & Nandi, A. K.
Computer Science, 10(3), 36-47. (2018). Credit Card Fraud Detection Using AdaBoost and Majority
[11] Khan, S., & Mishra, R. B. (2012). A neural network based approach for Voting. IEEE Access, 6, 14277–14284. doi:10.1109/access.2018.
English to Hindi machine translation. 2806420.
[12] Khan, S., Thirunavukkarasu, K., Hammad, R., Bali, V., & Qader, M. R. [23] Sahin, Y., Bulkan, S., & Duman, E. (2013). A cost-sensitive decision
(2021). Convolutional neural network based SARS-CoV-2 patients tree approach for fraud detection. Expert Systems with Applications,
detection model using CT images. International Journal of Intelligent 40(15), 5916-5923.
Engineering Informatics, 9(2), 211-228. [24] Shahnawaz, & Mishra, R. B. (2015). An English to Urdu translation
[13] Kumar, M. S., Soundarya, V., Kavitha, S., Keerthika, E. S., & Aswini, model based on CBR, ANN and translation rules. International Journal
E. (2019). Credit Card Fraud Detection Using Random Forest of Advanced Intelligence Paradigms, 7(1), 1-23.
Algorithm. 2019 3rd International Conference on Computing and [25] Shahnawaz, M. R. (2011). ANN and rule based model for English to
Communications Technologies (ICCCT). doi:10.1109/iccct2.2019 Urdu-Hindi machine translation system. In Proceedings of National
.8824930. Conference on Artificial Intelligence and agents: Theory& Application,
[14] Le Borgne, Yann-A., and Gianluca Bontempi. " Machine Learning for AIAIATA (pp. 115-121).
Credit Card Fraud Detection - Practical Handbook " Universit'e Libre de [26] Szmigiera, M. (2021). Value of fraudulent card transactions worldwide
Bruxelles, 2021, [Link] 2021-2027, [Link], accessed on 15-Jan-2022. [Link]
Machine_Learning_for_Credit_Card_Fraud_Detection_-_Practical_ .com/statistics/1264329/.
Handbook. [27] Xuan, S., Liu, G., Li, Z., Zheng, L., Wang, S., & Jiang, C. (2018).
[15] Lucas, Y., Portier, P. E., Laporte, L., He-Guelton, L., Caelen, O., Random forest for credit card fraud detection. 2018 IEEE 15th
Granitzer, M., & Calabretto, S. (2020). Towards automated feature International Conference on Networking, Sensing and Control (ICNSC).
engineering for credit card fraud detection using multi-perspective doi:10.1109/icnsc.2018.8361343.
HMMs. Future Generation Computer Systems, 102, 393-402. [28] Zhang, X., Han, Y., Xu, W., & Wang, Q. (2019). HOBA: A Novel
[16] Lusted, L. B. (1971). Signal detectability and medical decision-making. Feature Engineering Methodology for Credit Card Fraud Detection with
Science, 171(3977), 1217-1219. a Deep Learning Architecture. Information Sciences.
doi:10.1016/[Link].2019.05.023.

418 | P a g e
[Link]
View publication stats