AI Based Credit Card Fraud Detection using Machine

Learning Technique
Dr. B. Rajesh Kumar1
Department of Software Systems
Sri Krishna Arts and Science College

Sam Ashbal Raj J2

Department of Software Systems
Sri Krishna Arts and Science College

ABSTRACT have numerous ways to conceal their identity, so

As the number of companies and consumers
they can carry out their evil deeds in the warmth of
using credit card information for financial
their own homes. It is challenging to find those
transactions has grown, which has resulted in a
who utilise identity concealing techniques such as
substantial rise in fraud cases. This issue has
utilizing a VPN, passing the victim's
been made worse by dealing with outliers,
communication through the Tor network, etc. It is
unbalanced, and noisy data. This work proposes
critical to understand the consequences of monetary
the use of artificial intelligence for fraud
losses incurred via internet sources. Assuming card
detection. The proposed technique uses logistic
details are obtained, fraudsters have two options:
regression as a method for creating the classifier
either they use the payment cards for personal use
in order to avoid transactions using credit cards
or they sell the details of your card to various
fraud. Pre-processing ensures a high degree of
people, as is the case in India, where the card
precision in detection and is employed to control
information of around seventy million individuals
dirty data. The pre-processing stage of cleaning
is sold through the dark web. One of the biggest
the data uses two distinct basic techniques: the
credit card fraud incidents in UK history led to
mean-based methodology and the clustering-
GBP 17 million in financial damages. A group of
based strategies. The proposed classifier yields
international con artists came together in the
better results in terms of error rate, receptivity,
second decade of the the decade of 2000 to obtain
and accuracy.
the credit card information ofover 32,000 accounts.
The largest theft of credit cards in history is
Keywords : Credit cards, machine learning, and
thought to have occurred in this instance.
artificial intelligence
Thus, credit card fraud causes billion-
1. INTRODUCTION
dollar losses as a result of insufficient security
An increasing number of transactions will
measures. Card issuers are reassured that each
be made online as the globe moves toward a
transaction are benign, as are cardholders using
cashless future. Fraudsters nowadays do not need to
their cards and issuers executing the transactions.
be present at the scene of the crime to do it. They
On the other hand, fraudsters want to trick

209
2024, IRJEdT Volume: 06, Issue: 04 | April-2024
cardholders and financial institutions into thinking classifying process, especially when it came to
that the unauthorized transactions are real. terms of uncovering fraudulent transactions with
Furthermore, certain fraudulent transactions take credit cards.
place on a regular basis with the intention of The suggested method, which selects the
making money without the card issuers' or optimal machine learning algorithms first, uses a
cardholders' knowledge. The most disheartening variety of resampling strategies to highlight the
thing about payments made with credit cards is that imbalance class issue resulting from the dataset's
often neither cardholders nor the regulated uneven distribution of classes. This research takes
establishments realise they have been participating into consideration both improved cross-validation
in fraudulent behaviours. Because of this, it is often (CV) approaches and resampling strategies. This
exceedingly difficult to spot a fraudulent paper proposes a comprehensive way for choosing
transaction among thousands of genuine both the best machine learning algorithm and the
interactions, especially when there are significantly best reproducing strategy. This approach is based
fewer criminal transactions than lawful ones. The on a two-phase research that uses metrics to
banking and insurance industries can potentially be evaluate performance. Analyzing nine ML
kept safe from crime by utilising a variety of fraud techniques with their default settings is the goal of
detection techniques, such as predictive analytics the first step. The nine methods are: There are
and data mining, especially demonstrating several gradient booster machines: Using random
algorithms that employ clustering and anomaly forest (RF) algorithm, also known as the KNN
detection techniques. Machine learning algorithms algorithm), Decision Forest (DT), Naively Bayes
are necessary for most of these techniques, whether (NB), the extreme gradient booster (XGBoost), Lite
supervised or unsupervised, and they are helpful in Gradients boosters Machine (LightGBM), and
categorising credit card fraud. However, there are Classification Boosting (CatBoost). For usage in
several challenges that machine learning algorithms the second round, only the best three approaches
must overcome in order to detect every scam. An from the initial batch of nine are chosen. In the
ideal model based on machine learning would need second step, 19 distinct resampling techniques will
the highest values for common assessment metrics. be examined using all of the three methods that
To come closer to this perfect framework, this field were established in the previous phase. These 19
of research has to make a lot of breakthroughs. resampling methods fall into the following
Numerous components, such as the use of cross- categories: There are eleven undersampling, six
validation techniques for machine learning oversampling, and two that combine both
methods, and resampling strategies, influence the inadequate sampling and oversampling methods
issues associated with credit card fraud detection. simultaneously. In addition, the goal of this step is
These considerations can enhance the model's to determine which algorithm and reproducing
performance, as the evaluation metrics can attest. technique combination will yield the best suggested
In practical applications, balanced datasets are model in terms of overall performance. This
extremely uncommon, hence most of the period the innovative method stands out because it looks at
classification algorithm reduces the importance of many ways to address the issue of class imbalance
the minority class within the dataset. In reality, the in the dataset.
minority class played a critical role in the

210
2024, IRJEdT Volume: 06, Issue: 04 | April-2024
 This is illustrated by drawing present physically during physical transactions.
comparisons between the best machine learning Conversely, digital transactions happen over the
techniques, resampling, and using stratified K-fold phone or the internet. Typically, a cardholder gives
CV. Utilizing so many diverse approaches and their credit card numbers, card verification
strategies yields a hopeful outcome, especially in numbers, and expiration date over the phone or on
light of the actuality that gathering every single one a website. Credit card usage has skyrocketed due to
of the assessment metric values took more than a the recent explosive growth in e-commerce.
month. The remainder of the written piece was In Malaysia, there were roughly 317
organized in this manner. The second section offers million credit card interactions in 2011, and by
an overview of pertinent literature. In the third part, 2018, there were 447 million. According to, credit
the recommended approach is described. The final card theft hit a record $21.84 billion worldwide in
portion is going to address what was discovered 2015. With more people using credit cards, there
during the experiment. A synopsis of the results have been an increasing number of fraud
and recommendations for further study are occurrences. Despite the implementation of several
provided in the fifth section's conclusion. verification techniques, there has been little to no
According to the law, criminal activity is decline in credit card fraud instances. Fraudsters
defined as having the intent of misleading have a plethora of options due to the constantly
individuals with the purpose to benefit financially evolving financial services industry and the
or personally. Because of this, the two main possibility of significant financial advantages.
strategies for preventing loss from fraud are fraud Payment card fraud proceeds are frequently utilized
detection and prevention. While preventive fraud for illegal actions that are difficult to stop, such
prevention is the proactive strategy to prohibit financing terrorist attacks. Since they can hide their
fraudulent activities from arising, catching fraud is identity and location online, fraudsters tend to
the procedure of comprehending fraudulent gravitate towards it. The financial industry has been
transactions that are performed by fraudulently severely impacted by the recent rise in credit card
persons. These days, it's common to find a range of theft. Since merchants are responsible for all costs
payment cards, such as financing, charge, debit associated with credit card theft, including fees
cards, and prepaid cards. They're the most widely from card issuers, administrative costs, and other
used form of payment in certain nations. It's true penalties, they are primarily affected. The
that developments in technological advances have merchants bear full responsibility for any losses,
opened the door to adjustments in our financial which results in higher product pricing and lower
practices, particularly with regard to payment discounts. Therefore, minimising this loss is
methods, which have shifted from being physical crucial. Reducing the amount of fraudulent
actions to digital ones involving electronics. This transactions requires an efficient fraud detection
has completely changed the environment in which system.
monetary policy is implemented, as well as how 2.RELATED WORKS
big and small businesses operate. using a credit Credit card fraud breaks into two categories:

card without authorization to pay for goods or external fraudulent and internal fraud. 12, 15] Three

services. These types of transactions can be carried different groups have been set up for a more detailed
classification: Internet fraudsters (credit card generators,
out electronically or physically. The credit card is
however, website replication, and deceptive seller's

211
2024, IRJEdT Volume: 06, Issue: 04 | April-2024
pages), merchant-related forgeries (seller a conspiracy sampling failure scenarios and excessive sampling
and triangulation), and common account-related thefts instances that were positive is used in this study.
(realization, stolen, considered takeover, fake, and The basis for identifying credit card fraud is a
counterfeit). [16]. According to [17], there were over thorough examination from cardholder habits of
USD 16 billion in global fraud losses for banks and spending. This expenditure profile is examined using the
businesses in 2014. This represents a nearly 2.5 billion best combination of parameters in order to fully capture
United States dollars increase in losses from the previous the unique purchasing patterns of the card's user.
year. According to the study, there were instances of 5.6 Illegitimate transaction activity as well as legitimate
dollars of criminal activity associated with each $100. usually have dynamic characteristics. Therefore, the best
The main factor that sets apart credit card transaction characteristics need to be selected in a way that clearly
data is an unusual event. Typically, there are similar distinguishes the two profiles in order to categorize
features that set lawful and illegal transactions transactions made using credit cards efficiently. The
distinctive. Fraudsters pick up cutting-edge methods to effectiveness of fraud with credit card identification
imitate real cardholders' banking habits. As a result, the systems is influenced on the elements that comprise the
definition of what kind of conduct is honest or unethical history of card usage and the methods employed. These
continues to evolve. Due to this inherent feature, there variables are obtained by combining transaction data
are fewer genuine fraudulent cases found in a collection with a credit card's historical transaction history. Time-
of payment card activity information, which leads to based amount data, limited-time volume activities
distributions that have a strong bias in favor of the figures, retailer variety statistics, geographical
negative group (legal transactions). Twenty percent of information, and overall transactions statistics are the
the instances that are positive, 0.025% of the positive five main categories into which these variables may be
instances [19], and less than 0.005% of the favorable divided [19]. The card's total use profile is displayed by
cases [8] are included in the financial card information the variables that fall within the broad scope of every
examined in [18]. The data employed for the above transaction's information type. The cardholder's spending
research demonstrated that 0.172% of all of the habits are displayed by the variables under the locale
transactions belong to the advantageous category statistics type, which takes geographic regions into
(frauds). The severely distorted credit card data has been consideration. The factors listed underneath the
tested through a number of sampling procedures. [18, 20] merchant's statistics type reflect the debit card's usage in
employ a random sample technique to explain several merchant classifications. Each card's application
experimental findings that show classification algorithms character is defined by the parts of time-based statistics
with the lowest rates of false positives and the highest kinds in terms of their utilization values versus instant
true positive rate are produced when training information intervals or usage numbers versus time ranges. The
related to fraud as well as non-fraud is purposefully majority of publications emphasised cardholder profiles
distributed 50:50. above card profiles. It goes without saying that a single
In the publication [8], stratified sampling is person can make different types of transactions with
utilized to significantly undersample the legitimate many account numbers for credit cards.. As a result,
records. The experiment comparing the 50:50, 10:90, as using cardholders allows one to offer a distinct spending
well as 1:99 sample sizes show that the aforementioned profile. Since a cardholder may demonstrate a variety of
10:90 equivalent matches deception as well as legitimate behaviors across several cards on your own, whereas one
incidents. most satisfactorily on the basis of efficiency transaction on a card can only present a single spending
assessments on the 1:99 set since it is most comparable profile, the focus is this investigation is on the card's use
to the actual distributed that contains both legitimate and rather than the cardholder. Thirty variables are used in
fraudulent occurrences. In [21], a stratified sampling [19], (20 variables have been reduced to sixteen relevant
method is also performed. To maintain important trends ones.
from the data gathered, a combination of inadequate

212
2024, IRJEdT Volume: 06, Issue: 04 | April-2024
Fraud rates typically rise when credit cards surpass cash
as the most widely used form of payment for internet and
offline payments. With the introduction of big data, the
conventional human approaches to fraud detection are
growing more difficult due to their imprecise and time-
consuming nature. But these days, financial companies
are employing cunning tactics. Computerized intelligence
(CI) is the foundation of these intricate deception
techniques. Supervision and unsupervised techniques
make up the two primary groups of analytical fraud
detection methods [22]. In managed techniques for
authentication, abnormality transactions are recognized Figure 3.1 Overall Proposed Architecture
as possible instances of false actions when performing
unsupervised fraud identification, while models have 3.1 DATA SET DESCRIPTION
been built and computed upon samples of honest and The The University of Louisiana at ML
bogus deals [13] to categorize recent activities as
Group provided the dataset, and they also have an
dishonest or lawful. You may find an extensive review of
explanation. The dataset includes credit card
both supervised and unsupervised methods in [23].
purchases made in September 2013 by consumers
Identifying bogus credit card purchases has been the
focus of numerous studies employing various
across Europe. A total in 284,807 transactions over

methodologies. the course of two days are included in this dataset.

2.1 EXISTING SYSTEM The positive class, which makes up 0.172% of the
Furthermore, earlier systems had issues in real-time bank transaction data, is comprised of fraud events.
settings. These issues relate to the identification of credit There is a noticeable imbalance in the statistics,
card fraud. These issues include data that is unbalanced, skewed in favor of the healthier class. Its sole input
noisy, and pertains to drift. The bag construction method, variables are quantitative (continuous), obtained by
which comprises using the collected data to perform a
transforming a feature set using basic component
real-time procedure for sampling, was employed by the
analysis (PCA) and producing 28 principle
authors to address the data issues. They used logistic
components as a result. Therefore, a total of thirty
regression to effectively manipulate noisy data in order to
clean it up. To address the idea of drift, a strategy based
input characteristics are used in this investigation.

on incremental learning was suggested. Security issues hinder the specifications and
3. METHODOLOGY context of the features from being disclosed. The
seconds that pass between each trade and the
dataset's initial transaction are included in the time
feature. The total amount of the transaction is
shown by the 'amount' feature. When there is no
fraud present, the binary classification's targeting
class, feature "class," receives a value of 0, and
when there is fraud present, it receives a value of 1.
(a scam).
3.2 DATA PRE-PROCESSING
At this stage, the data needs to be cleaned up and
prepared for the classifier's training phase. Real-
213
2024, IRJEdT Volume: 06, Issue: 04 | April-2024
world data are typically noisy. As such, a cleaning 3.4 TRAIN THE MODEL
procedure is required. The steps involved in an
LOGISTIC REGRESSION
information cleansing process are as follows:
Using a functional approach, logistic
1The absence of digits should be entered. An input
regression calculates the likelihood of a binary
error caused a cell with an absent answer inside an
response. based on many factors (features). It
elected record to be empty..
ascertains which parameters best fit the sigmoid, a
2) Address any discrepancies. This implies that a
nonlinear function. The sign geometry function (σ)
data collision needs to be fixed if it occurs.
and the algorithm's input (x).
3) Remove any irregularities. Outliers are
quantities that are abnormal, such as levels that are
disproportionately high or low.
Fortunately, with the exception of a few missing
numbers, the majority of the data utilized in the set
of numbers are cleansed. and outliers.
Since the data are numerical, the method for
For every member of the the vector (z) that carries
addressing the missing values is based on the mean
the entered data, the optimal parameters (w) are
(a mathematical process). gives an illustration of
multiplied. to obtain the target class's classification
how to enter the missing value. To deal with
classification. The output of this operation is a
outliers, a clustering-based method is applied in
single integer. If a sigmoid's value is more than 0.5,
this work. The basic idea is to create three clusters:
it's considered a 1, else it's considered a 0. To find
one for the typical data, one for high computation,
the best-fit parameters, a classification system is
as well as one for low values. The final two
trained applying an optimization approach.
clusters—those with outliers—are eliminated once
Experiments were carried out to evaluate the
the data has been grouped into clusters. The
classifier's performance utilizing improved
Outlier Removal Mechanism.
randomly distributed gradient-descending
3.3 FEATURE EXTRACTION
optimization techniques and valley elevation (9).
We will distribute the goal value in this part, which
is essential for selecting relevant accuracy metrics
and, as a result, correctly evaluating various where the parameter ∆ represents the changing

machine learning models. Counting values of the magnitude of the incline ascent.. Until a halting
explanatory variable, often referred to as the condition is satisfied, the processes are repeated.

deciding variable, will be the first step in To find out if the criteria are converging, the

determining whether or not a transaction involving optimization techniques are examined throughout

credit card fraud is fraudulent. Secondly, we will iterations ranging from 50 to 1000. In other words,
distinguish between category and numerical do the parameters approach a fixed value or do they
features. Next, we will illustrate the relationship vary continuously? After 100 repetitions, steady
between the category features in a variety of plots variables are attained. Instead of updating the

and attempt to determine—or rather, observe—the classifier all at once when new information has
impact of those qualities on the actual determining been received, stochastic gradient ascent does so

variable, or “outcome.". gradually. At the beginning, the ratios are all set to

214
2024, IRJEdT Volume: 06, Issue: 04 | April-2024
1. The ascent from the gradient is then computed Step 4. Evaluation:
for each feature recurrence in the dataset. The Recall, precision, repeatability, and F1-score are
gradient is multiplied by alpha to update the among the metrics utilised to assess the model's
weights vector. The weight scalars is then returned performance. These measures evaluate the degree
after that. Because stochastic gradient ascent to which the projections made by the model and the
modifies weights one example at a time and is actual results agree.
computationally efficient, it is appropriate for the 3.5 TEST MODEL
large amount of data in our study. We test our framework's hypothesis using
ALGORITHM PROCEDURE this portion of the dataset. It remains unaltered and
Step 1. Model formulation: hidden until the equations and hyperparameters are
Describe the independent variables (elements determined. Only then is the model used to test data
influencing the outcome) and the variable that's to obtain a precise estimation of its performance
dependent (binary outcome you wish to predict). when applied to real-world data.
Select the sigmoid activation function (1 / (1 + 4. RESULT AND DISCUSSION
exp(-z))), for illustration. This function converts the
The calculation of precision involves
linear sum of characteristics and weights (z) to a
dividing the entire number of predictions made for
value (which indicates the likelihood that the event
a dataset by the entire amount of the correct
will occur) between 0 and 1.
predictions. It may immediately inform us of a
Step 2. Optimization:
model's training success as well as the overall
The algorithm minimizes a loss function (such as
performance trajectory that the model may follow.
log loss or cross-entropy) by iteratively optimizing
It does not, however, offer precise information on
the model parameters (weights). This quantifies the
how it relates to the issue. The accuracy, or a PPV,
difference between the actual result in binary terms
is an acceptable measure of outcomes even while
and the anticipated probabilities of the model.
rate of false positives are considerable. Turnover is
Gradient descent and other optimization techniques
the model measure that determines which model is
are used to gradually improve the model's ability to
best when one considers a large cost connected to
predict the binary decision by adjusting the weights
false negatives. Recall is beneficial, even when
in an area that minimizes the loss.
false negatives come at a high expense. To
Step 3. Prediction:
determine symmetry along recall and accuracy, you
The theory can be employed to perform prediction
need to get an F1-score.It works as a broad
once it has been trained. The model uses the
indicator of the degree of precision of the model. It
sigmoid function and weighted sum to calculate the
blends memory with accuracy. It is possible to
probability of an event happening for the latest
explain a high F1-score by having few negative
information with its attributes. The outcome is
results and few fraudulent positives.
generally characterized using a threshold value,
True Positives (TP):- These are the positive
usually 0.5; a forecast possibility exceeding the
numbers that were correctly predicted, indicating
threshold is categorized as positive (an event is
that both the real and projected outcomes for the
occurring), while a predicted probability below the
class are yes.
cutoff is labeled as negative (an event is not
True Negatives (TN):- These negative numbers
occurring).
demonstrate that the actual and predicted values for

215
2024, IRJEdT Volume: 06, Issue: 04 | April-2024
the class are both a value of zero proving thus the The number of accurately predicted groups that are
prediction was accurate. positive among all the positive samples in the
Values that appear when the predicted class as well dataset is referred to as "recall".
as our actual class diverge are known as fake Recall = TP/TP+FN
negatives and false positives
False Positives (FP):- When the class anticipated
is yes but the true classification is no, this is known
as a false positive (FP).
False Negatives (FN):- False Negatives (FN):
Arrangements for scheduled sessions where
instruction is not provided.

Figure 4.1 SCREENSHOT FOR DATA SET

Where,
 Correctly recognized (TP) means true
positive
 Erroneously detected as a false
Figure 4.2 Graph Representation for Fraudulent
positive (FP)
Transaction
 Correctly dismissed True Negative (TN)
 An inappropriate acceptance is usually
referred to as a false negative (FN)..
ACCURATE
Specifically, one must ascertain the proportion of
Forecasts for the positive subclass that accurately
represent the positive class.
Reliability = TP/TP+FP
RECALL

216
2024, IRJEdT Volume: 06, Issue: 04 | April-2024
 recommended classifier is evaluated using metrics
such as sensitiveness, accuracy, and the rate of
error. In comparison to the proposed logistic
regression-based classifier, two commonly used
classifiers are examined: the vote classifier and the
K-nearest peers classifier. The logistic regression-
based classifier, with accuracy of 97.2%,
responsiveness of 97%, and inaccuracy rate of
2.8%, yields the best results.
REFERENCES
[1]. Maes, S., Tuyls, K., Vanschoenwinkel, B. and
Manderick, B., (2022). Credit card fraud detection
Figure 4.3 Screen Shot Model training
using Bayesian and neural networks. Proceeding
International NAISO Congress on Neuro Fuzzy
Technologies.
[2]. Ogwueleka, F. N., (2021). Data Mining
Application in Credit Card Fraud Detection
System, Journal of Engineering Science and
Technology, Vol. 6, No. 3, pp. 311 – 322.
[3]. RamaKalyani, K. and UmaDevi, D., (2019).
Fraud Detection of Credit Card Payment System by
Genetic Algorithm,International Journal of
Scientific & Engineering Research, Vol. 3, Issue 7,
pp. 1 – 6, ISSN 2229-5518.

Figure 4.4 Screen Shot For Accuracy Score [4]. Meshram, P. L., and Bhanarkar, P., (2019).

5. CONCLUSION Credit and ATM Card Fraud Detection Using

The detection scientific credit card fraud is a Genetic Approach, International Journal of

crucial field of study. Financial institutions are Engineering Research & Technology (IJERT), Vol.

reporting a rise in fraud cases, which is the cause of 1 Issue 10, pp. 1 – 5, ISSN: 2278-0181.

this. This issue opens the door to developing fraud [5]. Singh, G., Gupta, R., Rastogi, A., Chandel, M.

detection systems with artificial intelligence. An D. S., and Riyaz, A., (2018). A Machine Learning

automated intelligence ( also known as AI Approach for Detection of Fraud based on SVM,

fraudulent activity detection system's machine International Journal of Scientific Engineering and

learning algorithm, also known as a classifier, Technology, VolumeNo.1, Issue No.3, pp. 194-

requires training from a database. Actually, there 198, ISSN : 2277-1581.

are outliers, noisy information, and values that are [6]. Seeja, K. R., and Zareapoor, M., (2014).

not present in the tainted data. The accuracy rate of FraudMiner: A Novel Credit Card Fraud Detection

the system is negatively impacted by these issues. Model Based on Frequent Itemset Mining, The

The above issues are thought to be resolved using a Scientific World Journal, Hindawi Publishing

logistic regression-based classifier. The

217
2024, IRJEdT Volume: 06, Issue: 04 | April-2024
Corporation, Volume 2018, Article ID 252797, pp. fraud: A comparative study. Decision Support
1 – 10, http://dx.doi.org/10.1155/2014/252797 Systems, 50(3), 602-613.

[7]. Patil, S., Somavanshi, H., Gaikwad, J.,

Deshmane, A., and Badgujar, R., (2018). Credit
Card Fraud Detection Using Decision Tree
Induction Algorithm, International Journal of
Computer Science and Mobile Computing
(IJCSMC), Vol.4, Issue 4, pp. 92-95, ISSN: 2320-
088X.
[8]. Duman, E., Buyukkaya, A., & Elikucuk, I.
(2019). A novel and successful credit card fraud
detection system implemented in a turkish bank. In
Data Mining Workshops (ICDMW), 2013 IEEE
13th International Conference on (pp. 162-171).
IEEE.
[9]. Bahnsen, A. C., Stojanovic, A., Aouada, D., &
Ottersten, B. (2019). Improving credit card fraud
detection with calibrated probabilities. In
Proceedings of the 2014 SIAM International
Conference on Data Mining (pp. 677-685). Society
for Industrial and Applied Mathematics.
[10]. Ng, A. Y., and Jordan, M. I., (2021). On
discriminative vs. generative classifiers: A
comparison of logistic regression and naive bayes.
Advances in neural information processing
systems, 2, 841-848.
[11]. Maes, S., Tuyls, K., Vanschoenwinkel, B., &
Manderick, B. (2022). Credit card fraud detection
using Bayesian and neural networks. In
Proceedings of the 1st international naiso congress
on neuro fuzzy technologies (pp. 261-270).
[12]. Shen, A., Tong, R., & Deng, Y. (2018).
Application of classification models on credit card
fraud detection. In Service Systems and Service
Management, 2007 International Conference on
(pp. 1-4). IEEE.
[13]. Bhattacharyya, S., Jha, S., Tharakunnel, K., &
Westland, J. C. (2019). Data mining for credit card

218
2024, IRJEdT Volume: 06, Issue: 04 | April-2024