Credit Card Fraud Detection Using Machine

Learning

by
Group:23

Md. Shakil Hossain (Exam Roll: 192340, Class Roll: 2023)

Mahabubur Rahman (Exam Roll: 192341, Class Roll: 2024)
Nahidul Islam (Exam Roll: 192345, Class Roll: 2028)

A Project Report submitted to the

Institute of Information Technology
in partial fulfillment of the requirements for the degree of
Bachelor of Science in Information and Communication Technology

Md. Mahmudur Rahman

Lecturer
Institute of Information Technology
Jahangirnagar University

Institute of Information Technology

Jahangirnagar University
Savar, Dhaka-1342
October, 2023
 DECLARATION

This Artificial Intelligence lab project report is submitted to the Institute of In-
formation Technology, Jahangirnagar University, Savar, Dhaka, to partially fulfil the
requirements for having the [Link]. (Hons.) degree in ICT. This is also needed to
certify that the project work is under the IIT’s 4th Year 1st Semester course ”ICT
4102: Artificial Intelligence Lab”. So, we declare that this project report has not
been submitted elsewhere for the requirement of any degree, diploma or publication.

[Link] Hossain
Roll:2023

Nahidul Islam
Roll:2028

Mahbubur Rahman
Roll:2024

ii
 ACKNOWLEDGEMENTS

We feel pleased to have the opportunity of expressing our heartfelt thanks and
gratitude to those who all rendered their cooperation in making this [Link]
thesis is performed under the supervision of Md. Mahmudur Rahman,Lecturer ,
Institute of Information Technology (IIT), Jahangirnagar University, Savar, Dhaka.
During the work, he has supplied us a number of books, journals, and materials
related to the present investigation. Without his help, kind support and generous
time spans he has given, we could not perform the project work successfully in due
time. First and foremost, we wish to acknowledge our profound and sincere gratitude
to him for his guidance, valuable suggestions, encouragement and cordial cooperation.
We express our gratitude to all other sources from where we have found help. We
are indebted to those who have helped us directly or indirectly in completing this
[Link] but not least, we would like to thank all the staff of IIT, Jahangirnagar
University and our friends who have helped us by giving their encouragement and
cooperation throughout the work.

iii
 ABSTRACT

Credit card fraud is a significant problem, with billions of dollars lost each year.
Machine learning can be used to detect credit card fraud by identifying patterns that
are indicative of fraudulent transactions. Credit card fraud refers to the physical
loss of a credit card or the loss of sensitive credit card information. Many machine-
learning algorithms can be used for detection. This project proposes to develop a
machine-learning model to detect credit card fraud. The model will be trained on a
dataset of historical credit card transactions and evaluated on a holdout dataset of
unseen transactions.
Keywords: Credit Card Fraud Detection, Fraud Detection, Fraudulent Transac-
tions, K- Nearest Neighbors, Support Vector Machine, Logistic Regression, Decision
Tree.

iv
 LIST OF ABBREVIATIONS

IIT Institute of Information Technology

JU Jahangirnagar University
ML Machine Learning
CCFD Credit Card Fraud Detection
KNN K- Nearest Neighbors
SVM Support Vector Machine
L.R. Logistic Regression
DT Decision Tree

v
 LIST OF FIGURES

Figure

4.1 Data Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

4.2 Data Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

4.3 Class Count . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

4.4 Data Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

4.5 Null data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

5.1 KNN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

5.2 Error Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

5.3 K=3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

5.4 K=7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

5.5 Logistic Regression Model . . . . . . . . . . . . . . . . . . . . . . . 16

5.6 Accuracy on Train data . . . . . . . . . . . . . . . . . . . . . . . . . 16

5.7 Accuracy on Test data . . . . . . . . . . . . . . . . . . . . . . . . . 17

5.8 Support Vector Machine Algorithm . . . . . . . . . . . . . . . . . . 17

5.9 SVM Confusion Matrix . . . . . . . . . . . . . . . . . . . . . . . . . 18

5.10 SVM ROC Curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

5.11 Decision Tree Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 19

5.12 D.T. Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

vi
5.13 Decision Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

5.14 Confusion Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

5.15 Table of Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

vii
 TABLE OF CONTENTS

DECLARATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii

ACKNOWLEDGEMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . iii

ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv

LIST OF ABBREVIATIONS . . . . . . . . . . . . . . . . . . . . . . . . . v

LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi

CHAPTER

I. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Project goals . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

II. Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . 3

III. Project Description . . . . . . . . . . . . . . . . . . . . . . . . . . 6

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.2 Data Source . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

IV. Data Analysis and Processing . . . . . . . . . . . . . . . . . . . 7

4.1 Data Description . . . . . . . . . . . . . . . . . . . . . . . . 7

4.2 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

V. Algorithm and Performance Analysis . . . . . . . . . . . . . . . 11

5.1 K-Nearest Neighbor (KNN): . . . . . . . . . . . . . . . . . . . 11

viii
 5.1.1 Algorithm KNN . . . . . . . . . . . . . . . . . . . . 11
5.2 Logistic Regression (L.R.): . . . . . . . . . . . . . . . . . . . 12
5.3 Support Vector Machine (SVM): . . . . . . . . . . . . . . . . 13
5.4 Decision Tree (D.T.): . . . . . . . . . . . . . . . . . . . . . . 14
5.5 Evaluation and Deployment . . . . . . . . . . . . . . . . . . . 15

VI. Future Work & Conclusion . . . . . . . . . . . . . . . . . . . . . 22

6.1 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

6.2 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

ix
 CHAPTER I

Introduction

1.1 Overview

With the increase of people using credit cards in their daily lives, credit card com-
panies should take special care of the security and safety of the customers. According
to (Credit card statistics 2021), the number of people using credit cards worldwide
was 2.8 billion in 2019; also, 70those users own a single card. Reports of Credit
card fraud in the U.S. rose by 44.7in 2020. There are two kinds of credit card fraud,
and the first is having a credit card account opened under your name by an identity
thief. Reports of this fraudulent behaviour increased 48to 2020. The second type is
when an identity thief uses an existing account you created, usually by stealing the
information on the credit card. Reports on this type of Fraud increased 9to 2020
(Daly, 2021). Those statistics caught We’s attention as the numbers have increased
drastically and rapidly throughout the years, which motivated We to resolve the issue
analytically by using different machine learning methods to detect fraudulent credit
card transactions within numerous transactions.

1.2 Project goals

The main aim of this project is the detection of fraudulent credit card transactions,
as it is essential to figure out the fraudulent transactions so that customers do not
get charged for the purchase of products that they did not buy. Fraudulent Credit
card transactions will be detected with multiple ML techniques. Then, a comparison
will be made between the outcomes and results of each method to find the best
and most suited model for detecting fraudulent credit card transactions; graphs and
numbers will also be provided. In addition, it explores previous literature and different
techniques used to distinguish Fraud within a dataset.

1
 Research question: What machine learning model is most suited for detecting
fraudulent credit card transactions?

2
 CHAPTER II

Literature Review

2.1 Introduction

Credit card companies must distinguish fraudulent from non-fraudulent transac-

tions so that their customers’ accounts will not get affected and charged for products
they did not buy (Maniraj et al.,2019). Many financial Companies and institutions
lose massive amounts of money because of Fraud and fraudsters that are seeking
different approaches continuously to violate the rules and commit illegal actions;
therefore, systems of fraud detection are essential for all banks that issue credit cards
to decrease their losses (Zareapoor et al., 2012). Multiple methods are used to detect
fraudulent behaviours, such as Neural Networks (N.N.), Decision Trees, K-nearest
neighbour algorithms, and Support Vector Machines (SVM). Those ML methods can
be applied independently or collectively with ensemble or meta-learning techniques
to develop classifiers (Zareapoor et al., 2012).

2.2 Literature Review

Zareapoor and his research team used multiple techniques to determine the best-
performing model for detecting fraudulent transactions, which was established using
the Accuracy of the model, the speed of detection and the cost. The models used
were Neural Network, Bayesian Network, SVM, KNN and. The comparison table in
theresearch paper showed that the Bayesian Network was high-speed in finding fraud-
ulent transactions with high Accuracy. The N.N. performed well, as the detection
was fast, with a medium accuracy. KNN’s speed was good with medium Accuracy,
andfinally, SVM scored one of the lower scores, as the speed was low, and the Accu-
racy wasmedium. As for the cost, All models built were expensive (Zareapoor et al.,

3
2012).[1]

The model used by Alenzi and Aljehane to detect Fraud in credit cards was Logis-
tic Regression. Their model scored 97.2% in Accuracy, 97% sensitivity and 2.8% Error
Rate. A comparison was performed between their model and two other classifiers ,
which are3 They were voting Classifier and KNN. V.C. scored 90% in Accuracy, 88%
sensitivity and 10% error rate, as for KNN where k = 1:10, the Accuracy of the model
was 93%, the sensitivity 94%and 7% for the error rate (Alenzi Aljehane, 2020).[2]

Maniraj’s team built a model to recognize if any new transaction is Fraud or non-
fraud. Their goal was to get 100% in detecting fraudulent transactions andtry to
minimize the incorrectly classified fraud instances. Their model has performed well
as they got 99.7% of the fraudulent transactions (Maniraj et al., 2019).[3]

The classification approach used by Dheepa and Dhanapal was the behaviour-
based classification approach, using a Support Vector Machine, where the behavioural
patternsof the customers were analyzed to distinguish credit card fraud, such as the
amount, date,time, place, and frequency of card usage. The Accuracy achieved by
their approach wasmore than 80% (Dheepa Dhanapal, 2012).[4]

Mailini and Pushpa proposed using KNN and Outlier detection in identifying
credit card fraud. After performing their model oversampled data, the authors found
that the most suitable method for detecting and determining target instance anomaly
is KNN, which showed that it is most suited to detecting Fraud with memory limita-
tion. As for Outlier detection, the computation and memory required for credit card
fraud detectionis much less in addition to working faster and better in large online
datasets. However, their work and results showed that KNN was more accurate and
efficient (Malini Pushpa,2017).[5]

Maes and his team proposed using Bayesian and Neural Networks to detect credit
card fraud. Their results showed that Bayesian performance is 8% more effective
in detecting Fraud than ANN, meaning BBN sometimes detects 8% more fraudu-
lent transactions. In addition to the Learning times, ANN can go up to several
hours,whereas BBN takes only 20 minutes (Maes et al., 2002).[6]

Jain’s team used several ML techniques to distinguish credit card fraud; three of

4
them are SVM, ANN and KNN. Then, to compare the outcome of each model, they
calculated the true positive (T.P.), false Negative (F.N.), false positive (F.P.), and
true negative (T.N.) generated. ANN scored 99.71% accuracy, 99.68% precision, and
0.12% false alarm [Link] accuracy is 94.65%, 85.45% for the precision, and 5.2%
false alarm rate. Moreover, finally,the Accuracy of KNN is 97.15%, the precision is
96.84%, and the false alarm rate is 2.88% (Jain et al., 2019).[7]

Dighe and his team used KNN, Logistic Regression and Neural Networks, multi-
layer perceptron and Decision Tree in their work, then evaluated the results regarding
numerous accuracy metrics. Of all the models created, the best performing one is
KNN, which scored 99.13%, then in second place performing model at 96.40% and in
last place is logistic Regression with 96.27% (Dighe et al., 2018).[8]

Sahin and Duman used four Support Vector Machine methods in detecting credit
card fraud. (SVM) Support Vector Machine with RBF, Polynomial, Sigmoid, and
Linear Kernel,all models scored 99.87% in the training model and 83.02% in the test-
ing part of the model (Sahin Duman, 2011).[9]

5
 CHAPTER III

Project Description

3.1 Introduction

In order to accomplish the objective and goal of the project, which is to find the
most suited model to detect credit card fraud, several steps need to be taken. Finding
the mostsuited data and preparing/preprocessing are the first and second steps; after
making surethat the data is ready, the modelling phase starts, where four models are
created: K-NearestNeighbor (KNN), Decision Tree, SVM and the last one is Logistic
Regression. In the KNN model, two Ks were chosen: K=3 and K =7. All models
were created in Jupiter Notebook programs

3.2 Data Source

The dataset was retrieved from an open-source website, [Link]. It con-

tains data on transactions made in 2013 by European credit card users in two days
only. Thedataset consists of 31 attributes and 284,808 rows. Twenty-eight attributes
are numeric variables that, due to the confidentiality and privacy of the customers,
have been transformed using PCA transformation; the three remaining attributes are
”Time”, which contains the elapsed seconds between the first and other transactions
of each Attribute, ”Amount” is the amount of each transaction, and the final attribute
“Class” which contains binary variableswhere “1” is a case of fraudulent transaction,
and “0” is not as case of fraudulent transaction.
Dataset : [Link]

6
 CHAPTER IV

Data Analysis and Processing

4.1 Data Description

The first figure below shows the structure of the dataset where all attributes
are shown, with their type, in addition to a glimpse of the variables within each
Attribute; as shown at the end of the figure, the Class type is integer, which I needed
to change to factor and identify the 0 as Not Fraud and the one as Fraud to ease
the process of creating the modeland obtain [Link] second figure shows
the class distribution; the red bar, which contains 284,315variables, represents the
non-fraudulent transactions, and the blue bar, with 492 variables, represents the
fraudulent transactions.

4.2 Analysis

7
Figure 4.1: Data Structure

Figure 4.2: Data Description

8
 Figure 4.3: Class Count

Figure 4.4: Data Correlation

9
Figure 4.5: Null data

10
 CHAPTER V

Algorithm and Performance Analysis

5.1 K-Nearest Neighbor (KNN):

This supervised learning technique achieves consistently high performance com-

pared to other fraud detection techniques of supervised statistical pattern recogni-
tion. Three factors majorly affect its performance distance to identify the least distant
neighbours. There are some rules to deduce a categorization from the k-nearest neigh-
bour and the count of neighbours to label the new sample. This algorithm classifies
transactions by computing the least distant point to this particular transaction. If this
least distant neighbour is classified as fraudulent, the latest marketing is also labelled
as fraudulent. Euclidean distance is an excellent choice to calculate the distances in
this scenario. This technique is fast and results in fault alerts. Its performance can
be improved by distance metric optimization.

5.1.1 Algorithm KNN

• Let m be the number of training data samples. Let p be an unknown point that
needs to be classified

• Storing the training samples in an array of data points arr[]. Each element of
this array denotes a tuple (x, y).

• for i = 0 to m do

• Calculating distance d(arr[i], p)

• end for

• They are making set S of K smallest distances achieved. Each of these distances
resembles an already classified data point.

11
 • Returning the majority label among S

Figure 5.1: KNN

Two Ks were used to determine the best KNN model, K=3 and K =7.
• K = 3 While making the KNN model, We created two models: K =3 and K =7.
Figure 5 shows the model created in Jupiter Notebook; the model scored an accuracy
of 100% and identified 85,443 transactions correctly and missed 131.
•K=7
There was a slight decrease in the Accuracy of the model created in Jupiter Note-
book (Figure 6) as it scored 100% when K is 7, and the model miss classified 131
fraudulent transactions as no fraudulent. As for the Accuracy is the same as K=3
100% with 52 misclassified transactions; the only difference is within the classifica-
tions.

5.2 Logistic Regression (L.R.):

This statistical classification model based on probabilities detects Fraud using a

logistic curve. Since the value of this logistic curve varies from 0 to 1, it can be

12
 Figure 5.2: Error Rate

used to interpret class membership probabilities. The dataset fed as input to the
model is classified for training and testing the model. Post-model activity is tested
for some minimum threshold cut-off value for prediction. Based on some threshold
probabilities, the logistic Regression can divide the plane using a single line and divide
dataset points into exactly two regions.
The last model created using Jupiter Notebook is Logistic Regression; the model
managed to score an Accuracy on Training data of 93.51% , while it scored an Accu-
racy score on Test Data of 91.88%, as presented in blew Figure.

5.3 Support Vector Machine (SVM):

Support vector machines or SVMs are linear classifiers, as stated in that work in
high dimensionality because, in high dimensions, a non-linear task in input becomes
linear. Hence, this makes SVMs highly useful for detecting Fraud. Its two most
important features that is a kernel function to represent the classification function
in the dot product of input data point projection and the fact that it tries finding
a hyperplane to maximize separation between classes while minimizing overfitting of
training data; it provides a very high generalization capability.
Finally, the model Support Vector Machine, as shown in Figure 12, scored 97.59%
for the Accuracy.

13
 Figure 5.3: K=3

5.4 Decision Tree (D.T.):

A supervised learning algorithm is a decision tree in the form of a tree structure

consisting of the root node and other nodes split in a binary or multi-split manner
further into child nodes, with each tree using its algorithm to perform the splitting
process. With the tree growing, there may be possibilities of overfitting the training
data with possible anomalies in branches, some errors or noise. Hence, pruning is
used for improving classification performance of the tree by removing specific nodes.
Ease in use and the flexibility that the decision trees provide to handle different data
types of attributes make them quite popular.
Algorithm: of Decision Tree

• Create (T)

• Calculate frequencies (Ci, T)

• If all instances belong to the same class, the returning leaf

• for every Attribute, a test is set for splitting criteria. An attribute that satisfies
the test is test node K

• Repeating Create (Ti) on each partition [Link] those nodes as children of

node K

14
 Figure 5.4: K=7

5.5 Evaluation and Deployment

The last stage of the CRISP-DM model is the evaluation and deployment stage,
as presented in Table 2 below. All models are being compared to determine the best
model for identifying fraudulent credit card transactions. Accuracy is the overall
number of instances that are predicted correctly; accuracies are represented by a
confusion matrix where it shows the True Positive (T.P.), True Negative (T.N.), False
Positive (F.P.) and False Negative (F.N.). True Positive represents the transactions
that are fraudulent and were correctly classified by the model as fraudulent. True
Negative represents the not fraudulent transactions that the model correctly predicted
as not fraudulent. The third rating is False positive, which represents the fraudulent
transaction but was misclassified as not fraudulent. Moreover, finally, False Negative,
which are the not fraudulent transactions identified as fraudulent; Table 1 below
shows the confusion matrix.
The table above shows all the components to calculate the Accuracy of a model,
which is displayed in the below equation.
Table 2 shows all of the accuracies of all the models that were created in the
project; all models performed well in detecting fraudulent transactions and managed
to score high accuracies. Out of all the models, the model that scored the best is
KNN and Decision Tree as its Accuracy is 100%, the third place is the Support Vector

15
 Figure 5.5: Logistic Regression Model

Figure 5.6: Accuracy on Train data

Machine, and the model that scored the lowest Accuracy out of all models is Logistic
Regression with a score of 93.51%.

16
 Figure 5.7: Accuracy on Test data

Figure 5.8: Support Vector Machine Algorithm

17
Figure 5.9: SVM Confusion Matrix

Figure 5.10: SVM ROC Curve

18
Figure 5.11: Decision Tree Algorithm

Figure 5.12: D.T. Accuracy

19
 Figure 5.13: Decision Tree

Figure 5.14: Confusion Matrix

20
Figure 5.15: Table of Accuracy

21
 CHAPTER VI

Future Work & Conclusion

6.1 Future Work

There are many ways to improve the model, such as using it on different datasets
with various sizes and data types or by changing the data splitting ratio and viewing
it from a different algorithm perspective. An example can be merging telecom datato
calculate the location of people to have better knowledge of the location of the card
owner while his/her credit card is being used; this will ease the detection because if
the card owner is in Dubai and a transaction of his card was made in Abu Dhabi, it
will easilybe detected as Fraud..

6.2 Conclusion

In conclusion, the main objective of this project was to find the most suited model
for creditcard fraud detection in terms of the machine learning techniques chosen for
the project. It was met by building the four models and finding the accuracies of
them all; the bestmodel in terms of accuracies is KNN and Decision Tree, which
scored 100the amount of credit card fraud and increase the customer’s satisfaction as
it will provide themwith a better experience and feeling secure.

22
References

[1] Z. et al, “Analysis on credit card fraud detection techniques: Based on cer-
tain design criteria.” [Link]
[Link], 2012. Accessed: 26-oct-2023.

[2] . A. N. O. Alenzi, H. Z., “Fraud detection in credit cards using logistic re-
gression.” [Link]
12&Code=IJACSA&SerialNo=65, 2020. Accessed: 26-oct-2023.

[3] S. A. A. S. . S. S. D. Maniraj, S. P., “Credit card fraud detection using machine

learning and data science.” [Link]
2019. Accessed: 25-oct-2023.

[4] . D. R. Dheepa, V., “Analysis on credit card fraud detection techniques: Based
on certain design behaviour-based credit card fraud detection using support vec-
tor machinescriteria.” [Link] 2012. Ac-
cessed: 26-oct-2023.

[5] . P. M. Malini, N., “Analysis of credit card fraud identification techniques based on
knn and outlier detection.” [Link]
2017. Accessed: 26-oct-2023.

[6] T. K. V. B. . M. B. Maes, S., “Credit card fraud detection using

bayesian and neural networks.” [Link]
credit-card-fraud-detection-using-machine-learning-and-data-science-IJERTV8IS09
pdf, 2002. Accessed: 23-oct-2023.

[7] N. S. . J. S. Jain, Y., “A comparative analysis of various credit card fraud detection
techniques,” 2019. Accessed: 25-oct-2023.

[8] P. S. . K. S. Dighe, D., “Detection of credit card fraud transactions using machine
learning algorithms and neural networks.” [Link]
2018.8697799, 2018. Accessed: 26-oct-2023.

23
[9] . D. E. . Sahin, Y., “Detecting credit card fraud by decision trees and support
vector machines,” 2011. Accessed: 23-oct-2023.

24