Credit Card Fraud Detection Using Machine Learning

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 10 Issue: 05 | May 2023 www.irjet.net p-ISSN: 2395-0072
© 2023, IRJET | Impact Factor value: 8.226 | ISO 9001:2008 Certified Journal | Page 1478
Credit Card Fraud Detection Using Machine Learning
Zainab Firdous1, Sushma V2, Aftab Pasha S3, M Shahista Banu4, Najmusher H5
1,2,3,4Dept. of CSE, HKBK College of Engineering, Bangalore
5Professor, Dept. of CSE, HKBK College of Engineering Bangalore
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract – Fraud is the act of depriving a
person/organization of money through willingness,
deception or other unfair means. The unforeseen event of
Covid-19 has led to many people embracing digital
transactions and online shopping. This combinedwithother
benefits provided by credit card issuers such as rewards has
increased the usage of credit card, and in turn increased
credit card frauds. Credit card default can have serious
implications on credit card holder and can affect financial
stability of credit card issuers. There is a need for a system
that can predict defaults ahead of time so that appropriate
measures can be taken by credit card issuers. In this paper,
we have provided a comparative analysis of various
machine-learning algorithms often used in fraud detection
such as logistic regression, decision tree classifier, random
forest classifier and support vector machine classifier. The
models were compared on the basis of precision, recall and
accuracy to find the best algorithm for predicting probable
defaulters.
Key Words: Credit Card fraud, Credit Card default,
Machine Learning, Support Vector Machine, Decision
Tree, Logistic Regression, Random Forest
1.INTRODUCTION
The banks earn money by various means such as lending
loans to other customers using the depositor’s money. The
interest gained is the profit earnedbythebank,butwhen the
borrower defaults, the loan becomes a NPA and is a huge
blow to the bank’s statements. It was estimatedthatthetotal
amount of NPAs in India increased from 2.39 lakh crore in
2014 to 10.36 lakh crore in 2018. Therefore, an effective
system must be developed to curbthesedefaultseven before
they occur. Different types of defaults in finance include:
 Loan default: This is the most common type of
default. In this case the loan borrower fails to repay
the loan.
 Credit card default: This occurs when the creditcard
holder uses his/her credit card to buy items that
they cannot afford but doesn’t repay the money
spent.
 Bond default: when the organization/government
fails to repay the loan/principal amount, it is
considered as bond default.
We have focused on developing a system to predict credit
card defaulters. A comparative analysis was done on the
performance of various algorithms such as logistic
regression, decision tree, random forest and support vector
machine based on precision, recall and accuracy. The
traditional systems available such as CIBIL score uses
demographic data, credit history and so on to calculate a
score. This score is employed by many money lending
organizations to judge whether to issue the loan to this
applicant or not and to set the credit limit in case of credit
cards. However, this system only gives an idea to banks of
the level of risk involved in granting the person loan. Hence,
we have developed a model that predict with precision the
probability of a user defaulting.
2. LITERATURE REVIEW
The paper by author Yue Yu, [1] concentrates on the
importance of credit card default prediction by highlighting
the need of timely and precisepredictiontopreventfinancial
losses for both banks and the users,thenproceedingtomake
an outline of various machine-learning based algorithms by
explaining their principles, advantages and comparing their
performances. Yue Yu then introduces the dataset used in
the paper, which includes thirty thousand entries of credit
card users from a Taiwanese bank containing their previous
creditcardtransactioninformation,card-userdemographics,
and payment behavior to train and evaluate these
algorithms. The paper then presents an analysis of applying
the selected machine learning algorithms to the dataset and
evaluating their performance using parameters such as
accuracy, precision, recall, score of harmonic mean and
Receiver Operating Characteristic graph. The results show
that all the algorithms considered for the test produce good
results but artificial neural networks and support vector
machine giving better accuracy and performance rates. The
article ends by admitting some drawbacks and potential
scope for future study.
Authors, Yashna Sayjadah, Khairl Azhar Kasmiran, Ibrahim
Abaker Targio Hashem, and Faiz Alotaibi, [2] in their paper,
have recognized challenges associated with credit card
default prediction, such as imbalanced datasets and the
requirement for precise models. To analyze their study, the
authors obtained a dataset containing credit card
information and default status which is pre-processed by
handling missing values, encoding certain variables, and
conducting feature ascending.Followingwhichthedatasetis
categorized into training and testing purposes. Accuracy of
different algorithmswasnotedforthepossibleoccurrenceof
default credit card. Algorithms used here are logistic

regression, decision tree and random forest, the accuracy of
these were 75%, 64% and77%respectively.Baseduponthis
the authors have concludedthatbankscanadapttherandom
forest algorithm to detect the credit card defaulters so that
they can analyze the risk before giving the credit card to the
clients. In general, the paper captures a thorough
examination of credit card default prediction using machine
learning methods and a seemingly bright platformforfuture
hypothesis.
Alžbeta Bačová and František Babič’s paper, [3] draws
attention towards creating a model that can predict the
possibility of a customer bound to miss his payment. For the
purpose of their investigation, the authors employed a
dataset that contained generic data on credit card users.
From the thesis, it was deducedthatpredictiveanalyticsmay
effectively identify clients who are more likely to forget to
pay using a credit card. The researchers found that specific
variable feature, definitely influence howa predictioncanbe
made. The trained model was accurate, which hints that
there is a way for them to find purpose in the credit card
market. The paper establishes how predictive analyticsmay
be used to identify credit card defaulters and reduce
financial risks for credit card issuers. To sum it up, thepaper
serves as a valuable resource for credit card companies and
industry professionals keen in implementing predictive
analytics to predict defaults of credit card holders.
The writers Talha Mahboob Alam, Kamran Shaukat,Ibrahim
A. Hameed, Suhuai Luo, Muhammad Umer Sarwar, Shakir
Shabbir, Jiaming Li, and Matloob Khushi, [4] investigatesthe
problem of credit card default prediction using imbalanced
datasets. The authors start with citing the need of a proper
detection and prediction system to identify precise and
accurate default predicting project, as it helps
establishments identify dicey borrowers and make
calculated decisions regarding credit approvals. There is
always an imbalance present while going through different
data sets. This can lead to biased modelsthatperformpoorly
in predicting defaults. the impact of datasets on credit card
default prediction is investigated along with different
techniques to address this issue. The authors find that
regular classifiers trained on imbalanced datasets tend to
benefit the majority class,resultinginsmallerrecall scorefor
the defaults. However, by applying sampling techniques,the
performance of the classifiersimprovesdrastically.Adaptive
Synthetic Sampling consistently beats other techniques,
providing better prediction accuracy and recall precisionfor
default instances. The authors highlight the need for more
comprehensive and diverse credit card default datasets to
validate the findings of this study and to facilitate the
development of better models.
The paper prepared by authors,Sara Makki,ZainabAssaghir,
Yehia Taher, Rafiqul Haque, Mohand-Saïd Hacid,andHassan
Zeineddine, [5] introduces different approaches for
detecting credit card frauds occurring through imbalanced
datasets. The experimental arrangement involves using a
practical credit card transaction.Theauthorsnotethatwhile
hybrid sampling andcertain algorithmsprovidetherequired
results, choosing the proper algorithm may depend
invariably on the problem at hand. The paper shows
different overview of researches of existing approaches and
techniques to check the ones who have defaulted. They
arrive at a solution that no algorithm alone is suffice to
address the issue of imbalance in datasets. The research
work briefs about confusion matrix and related results are
formulated in tables. The confusion matrix, is also given by
the term error matrix, it is a table that is frequently used to
assess how well a categorization model is working. It
summarises the model's predictionsandthedegreetowhich
they agree with the actual labels or dataclasses.Thegenuine
class labels and anticipated class labels are represented,
respectively, by rows and columns in a grid that represents
the confusion matrix. For the purpose of detecting credit
card fraud, the paper offers an experimental investigation of
unbalanced classification algorithms. To increase the
accuracy of fraud detection,itevaluatesthebodyofresearch,
discusses various methodologies, shows experimental
findings, and stresses on balancing classes.
D. Tanouz, G V Parameswara Reddy, R Raja Subramanian, A.
Ranjith Kumar, D. Eswar, CH V N M Praneeth, [6] have made
a report about the growing threat of credit card fraud and
the necessity for effective detection systems. The approach
depicts promising results and contributes to the
development of strong fraud detection systems by
recognizing the rising threat of credit cardfraud,theauthors
dwell into a broader spectrum of the research of various
algorithms suitable for fraud detection. To validate their
proposed approach, the researchers conduct experiments
using a practical credit card dataset. They as well assess and
compare the parameters of the different algorithms
employed. The efficiency of machine learning algorithms in
detecting credit card fraud by leveraging patterns is marked
and regular techniques process the results and upon
tabulating the outcomes various features were considered
which were put into tables evident in the paper. Theauthors
examine approaches such as primary component study,
feature ranking and engineering, providingvisionsintotheir
impact on the accuracy and efficiency of the systems. The
authors also discuss the importance of dataset pre-
processing and handling techniques to address the inherent
class imbalance problem in credit card fraud datasets. The
authors conclude by making a reference to a number of
studies, research papers, and publications in the area of
credit card fraud detection.
3. METHODOLOGY
3.1 DATA
The dataset used for building the models was obtained from
UCI Machine Learning Repository. It consists of 30,000

observations and 24 features. This information was collected
from Taiwan as part of a research in the year 2005. It consists
of demographic details of the user, payment history, bill
amount and the amount paid by the user.
Feature Description
ID
It is a number used to identify
the user
LIMIT_BAL It is the limit of the credit card
SEX
Indicates the gender of the
user (1=male, 2=female)
EDUCATION
It indicates the level of
education of the user (1 =
graduate school, 2 =
university, 3 = high school,
4,5,6,0 = others)
MARRIAGE
It indicates the marital status
of the user (1=married, 2=
single, 3,0=other)
PAY_0 to Pay_6
It indicates the payment
status of the credit card from
April to September 2005(-2 =
no consumption, -1 = pay duly,
1 = one month delay, 2 = two-
month delay….9 = nine-month
delay)
BILL_AMT1 to
BILL_AMT6
It indicated the bill amounted
on the credit card from April
to September 2005
PAY_AMT1 to
PAY_AMT6
It indicates the amount paid
by the user
DEFAULT
It indicates if the user
defaulted next month or not
(0 = non-defaulter, 1 =
defaulter)
Table 1: Description of the dataset
3.2 EXPLORATORY DATA ANALYSIS
We have used matplotlib and seaborn libraries to visualize
and get better understanding of the data. It is clear from the
bar graph in figure-1 that the data is highly imbalanced i.e.,
the number of non-defaulters is far more than defaulters,
this can lead to a bias towards class 0(non-default) while
modelling.
Figure 1: Target variable visualization
The contribution of various categorical valuesthatisgender,
education and marital status was analyzed in terms of
default.
Figure 2: Visualizing gender in terms of default
It can be observed from figure-2 that there was generally
more female data in the dataset compared to male data, and
as a consequence more females have defaulted as compared
to males.

Figure 3: Visualizing educational level in terms of default
Majority of the users in the dataset have educational level of
university, the users with educational level have defaulted
more, followed by educational level of graduate and finally
high school.
Figure 4: Visualizing marital status in terms of default
It can be noticed from figure-4 that the dataset mostly has
users who are single. There is not much difference in
number of defaulters who are single and defaulter who are
married. A slight percent of ‘other’ category has also
defaulted.
Finally, a correlationheatmapwasplottedtounderstand
the relationship of various attributed with each other and
the target variable.
Figure 5: Correlation heatmap
It can be noticed that default is only positively correlated
with payment status features. The limit_bal is positively
correlated with bill_amt and pay_amt, whichisobvioussince
the credit card issuers tend to give higher credit limit to
applicants with good credit history. Also bill_Amt and
payment status features are also positively correlated to
each other and bill_amt and pay_Amt are also positively
correlated.
3.3 DATA PRE-PROCESSING
From the figures-3 and figure-4 it canbe noticedthatthereis
inconsistency in the labels of these categorical features, that
is multiple labels seem to be indicating the samecategory. In
‘Education’ categorical feature, the label 4,5,6 and 0 all
indicate the ‘other’ category. In ‘marriage’ categorical
feature, the label 3 and 0 indicate users in ‘other’ category.
This can cause confusion for the model in prediction of
default, so we have removed these inconsistencies by
combining these labels into one.
Some algorithms can be sensitive to categorical features.
Thus, we encoded these categorical features using the Label
Encoder class of scikit-learn library.
We dropped the feature ‘ID’ since it is only used to indicate
the user and has no contribution towards the prediction of
default. Then the dataset was split into the train set and the
test set (80:20) using the train_test_split class of scikit-learn
library.
Finally, we standardized the dataset by scaling the data in
terms of median absolute deviation using sklearn library.
3.4 MACHINE LEARNING ALGORITHMS
The main objective of the project is to predict probable
defaulters for which we have used four most popular
machine learning algorithms often used for fraud detection.
These are:

Logistic regression:
This is popular classification algorithm that can takes any
input value and gives the probability of the dependent
variable between 0 and 1. It then fits a sigmoid function and
when the new input data point comes it calculates its
probability, if the probability is greater than the threshold it
is predicted as 1 or it is predicted as 0.
Decision tree:
Decision tree builds tree that predicts the outcome of the
dependent variable. The model asks a questionandbasedon
the answer splits the dataset into subset, here one node
corresponds to the answer yes and another node
corresponds to answer no. The model keeps splitting until
max_depth value has reached or the node contains values
from same class.
Random forest:
Random forest is based on ensemble learning which
combines multiple decision tree and predicts dependent
variable by taking average of all the decision tree. This gives
better accuracy and reduces the problem of overfitting.
Support vector machine:
Support vector machine algorithm creates a decision
boundary (hyperplane) that separates default and non-
default cases. This hyperplane is chosen such that it
maximizes the margin between the two classes. The margin
is created by calculating the distance between the
hyperplane and support vectors.
K-Nearest Neighbor:
KNN algorithm stores all the data during the training phase
and when the new data point comes in, it calculates the
distance between the new point and the neighbors. Then
selects the stated number of nearest neighbors.Finally,the
number of default and non-default outcomes are counted,
whichever has the majority is assigned as theoutcomeof the
new data point’s dependent variable.
4. RESULTS
The various algorithms were evaluated on the basis of
accuracy, precision and recall. Accuracy is the measure of
predictions that are correctly predicted out of the total
predictions. Precision is measure of true positive that were
actually correct. Recall is measureoftruepositivesthatwere
correctly identified. High recall will lead to fewer defaulters
being precited as non-defaulters where as High precision
will reduce the number of defaulters predicted as non-
default. However, there is a trade-off between precision and
recall. For our problem statement it is important to have
high precision, since we don’t want our system to be
predicting non-defaulters as defaulter.
Algorithm Accuracy
KNN 80
Logistic regression 81
Decision tree 73
Random forest 81
Support vector machine 82
Table 2: Accuracy results
Decision tree has the least accuracy and the best accuracy
is given by SVM.
Algorithm
Precision
for class 0
Precision
for class 1
KNN 0.92 0.36
Logistic
regression
0.97 0.24
Decision tree 0.81 0.41
Random forest 0.94 0.36
SVM 0.96 0.36
SVM (changed
threshold)
0.84 0.64
Table 3: Precision results
It can be noticed that generally the precision is more for
class 0 (non-defaulter), this is due to the imbalance in
dataset. As a result, even though the accuracy is good for
some of the algorithms, they hardly can predict defaulters
correctly. But on changing the decision threshold of support
vector machine, the precision for class 1(defaulter)
drastically increased, which means it can correctly predict
64% defaulters and 84% non-defaulters.
Algorithm Recall for
class 0
Recall for
class 1
KNN 0.84 0.55
Logistic
regression
0.82 0.69
Decision tree 0.83 0.38
Random forest 0.84 0.63
SVM 0.83 0.70
SVM (changed
threshold)
0.94 0.37
Table 4: Recall results

In case of recall too, generally the recall for class 0 is more
than class 1. The best recall for class 1 was from support
vector machine. The confusion matrix for the various
algorithms were as follows:
Figure 6: KNN confusion matrix
Although KNN has high true positive and true negatives,still
the number of false positive and false negative is quite high.
Figure 7: Logistic regression confusion matrix
Logistic regression has high true positives, it performs quite
bad at predicting defaulters. Thus,logistic regressiondoesn’t
serve the purpose of this proposed system.
Figure 8: Decision tree confusion matrix
Even though decisiontreehaspredictedthegreatest number
of defaults by far, the true positives have reduced and false
negatives have increased, which means quite a lot of
defaulters were let off the radar.
Figure 9: SVM confusion matrix
SVM seems to be doing well with good number of true
positives and true negatives and low false negatives.
Figure 10: SVM changed threshold
However, changing the threshold to increase the precision
for class 1, provides far better results. The number of true
negatives has increased and the number of false positives
has decreased.
Figure 11: Random Forest confusion matrix
Random forest too has performed good with high true
positives, true negatives.
However, SVM with changed threshold seems to be
performing the best with both true positive and true
negatives being high and reducing false positives.

5 CONCLUSIONS
The main objective of this proposed system is to correctly
predict defaulters. We employed KNN, logistic regression,
decision tree, random forest and support vector machine to
find the best algorithm for the problem statement. We
evaluated these algorithms based of accuracy, precision,
recall and comparing their confusion matrix. The major
challenge we faced was the imbalanced dataset, however
under sampling or oversampling seems to be
misrepresenting the relationship betweenXand Yleading to
poor performance. Thus, we have used the dataset without
any sampling for modelling. It can be noticed that accuracy
alone cannot be used to judge the performance of a model.
Even though logistic regression has high accuracy, it
performed poorly on test set. In conclusion, both support
vector machine and random forest have good accuracy.SVM
with slightly changed threshold performs the best with
reduced number of false negatives and better recall and
precision. However, it is hard to narrow down on one of the
models, because of the limited user data available due to
confidentiality.
REFERENCES
[1] Yue Yu, “The Application of Machine Learning
Algorithms in Credit Card Default Prediction”,
International Conference on Computing and Data
Science, 978-1-7281-7106-7/20 © 2020 IEEE
DOI:10.1109/CDS 49703.2020.0050.
[2] Yashna Sayjadah, Khairl Azhar Kasmiran, Ibrahim
Abaker Targio Hashem, Faiz Alotaibi, “Credit Card
Default Prediction Using Machine Learning
Techniques”, 978-1-5386-7167-2/18 © 2018 IEEE.
[3] Alžbeta Bačová, František Babič, “Predictive Analysis
for Default of Credit Card Clients”, SAMI 2021, IEEE
19th World Symposium on Applied Machine
Intelligence and Informatics, January 21–23, Herl’any,
Slovakia, 978-1-7281-8053-3/21/$31.00©2021IEEE.
[4] Talha Mahboob Alam, Kamran Shaukat, Ibrahim A.
Hameed, Suhuai Luo, Muhammad Umer Sarwar, Shakir
Shabbir, Jiaming Li, Matloob Khushi, “An Investigation
Of Credit Card Default Prediction In The Imbalanced
Datasets”, DOI: 10.1109/ACCESS.2020.3033784.
[5] Sara Makki, Zainab Assaghir, Yehia Taher, Rafiqul
Haque, Mohand-Saïd Hacid, Hassan Zeineddine, “An
Experimental Study With Imbalanced Classification
Approaches for Credit Card Fraud Detection”, DOI:
10.1109/ACCESS.2019.2927226.
[6] D. Tanouz, G V Parameswara Reddy, R Raja
Subramanian, A. Ranjith Kumar, D. Eswar, CH V N M
Praneeth, “Credit Card Fraud Detection Using Machine
Learning”, Fifth International ConferenceonIntelligent
Computing and Control Systems ICICCS 2021, IEEE
Xplore Part Number CFP21K74-ART; ISBN: 978-0-
7381-1327-2.

Credit Card Fraud Detection Using Machine Learning

More Related Content

Similar to Credit Card Fraud Detection Using Machine Learning (20)

More from IRJET Journal (20)

Recently uploaded (20)

Credit Card Fraud Detection Using Machine Learning