Proceedings of the International Conference on Intelligent Computing and Control Systems (ICICCS 2020)

IEEE Xplore Part Number:CFP20K74-ART; ISBN: 978-1-7281-4876-2

Credit Card Fraud Detection Using

Machine Learning
Ruttala Sailusha V. Gnaneswar
Department Of Information Technology Department Of Information Technology
Velagapudi Ramakrishna Siddhartha Engineering College Velagapudi Ramakrishna Siddhartha Engineering College
Vijayawada, India Vijayawada, India
rsailusha99@gmail.com v.gnaneswar123@gmail.com

R. Ramesh G. Ramakoteswara Rao

Department Of Information Technology Department Of Information Technology
Velagapudi Ramakrishna Siddhartha Engineering College Velagapudi Ramakrishna Siddhartha Engineering College
Vijayawada, India Vijayawada, India
rameshraparla59@gmail.com grkraoganga@gmail.com

Abstract—Credit card fraud detection is presently the most

A credit card is said to be a fraud when some other person
frequently occurring problem in the present world. This is due to
the rise in both online transactions and e-commerce platforms. uses your credit card instead of you without your
Credit card fraud generally happens when the card was stolen authorization. Fraudsters steal the credit card PIN or the
for any of the unauthorized purposes or even when the fraudster account details to perform any of the unauthorized transactions
uses the credit card information for his use. In the present world, without stealing the original physical card. Using the credit
we are facing a lot of credit card problems. To detect the card fraud detection we could find out whether the new
fraudulent activities the credit card fraud detection system was transactions are fraud one or a genuine one.
introduced. This project aims to focus mainly on machine
learning algorithms. The algorithms used are random forest
algorithm and the Adaboost algorithm. The results of the two
algorithms are based on accuracy, precision, recall , and F1-
score. The ROC curve is plotted based on the confusion matrix.
The Random Forest and the Adaboost algorithms are compared
and the algorithm that has the greatest accuracy, precision, recall
, and F1-score is considered as the best algorithm that is used to
detect the fraud.

Keywords—credit card fraud, fraudulent activities, Random Figure1.Growth of Internet users[2]

Forest, Adaboost, ROC curve
The fraud that is committed may involve the card such as a
credit card or debit card. In this , the card itself acts as a
I. INTRODUCTION fraudulent source in the transaction. The purpose of
Credit card fraud is a growing concern in the present world committing the crime may be to obtain the goods without
with the growing fraud in the government offices, corporate paying money or to obtain the unauthorized fund. Credit cards
industries, finance industries, and many other organizations. are a nice target for fraud. The reason is that in a very short
In the present world, the high dependency on the internet is time a lot of money can be earned without taking many risks
the reason for an increased rate of credit card fraud and even the crime will take many weeks to be detected.
transactions but the fraud has increased not only online but
also offline transactions. Though the data mining
techniques[6] are used the result is not much accurate to
detect these credit card frauds. The only way to minimize
these losses is the detection of the fraud using efficient
algorithms which is a promising way to reduce the credit card
frauds. As the use of the internet is increasing[Figure.1], a
credit card is issued by the finance company. Having a credit
card means that we can borrow the funds. The funds can be
used for any of the purposes. When coming to the issuance
of the card, the condition involved is that the cardholder will
pay back the original amount they borrowed along with the
Figure.2 Growth of E-Commerce sites[9]
additional charges they agreed to pay.
As the use of the internet nowadays [Figure.2] is very commerce sites is increasing and thereby there is a huge chance
much increasing there may be many chances for the fraudsters of credit card fraud. So to avoid such credit card frauds, we
to commit the credit card frauds. The main fraud cases that are need to find out the best algorithm that reduces credit card
ongoing in the present world are in those of the e-commerce frauds.
sites. In the present generation, people are showing much
interest in getting things online rather than going and II. RELATED WORK
purchasing them, and due to this , the growth of the e-
978-1-7281-4876-2/20/$31.00 ©2020 IEEE 1
Authorized licensed use limited to: University of Exeter. Downloaded on June 28,2020 at 07:04:27 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Intelligent Computing and Control Systems (ICICCS 2020)
IEEE Xplore Part Number:CFP20K74-ART; ISBN: 978-1-7281-4876-2

New methods for credit card fraud detection with a lot of

Regression, J48, and Adaboost. Naïve Bayes on among the
research methods and several fraud detection techniques with a
classification algorithm. This algorithm depends upon Bayes
special interest in the neural networks, data mining, and
distributed data mining. Many other techniques are used to theorem. Bayes's theorem finds the probability of an event
detect such credit card fraud. When done the literature survey that is occurring is given. The Logistic regression algorithm is
on various methods of credit card fraud detection, we can similar to the linear regression algorithm. The linear
conclude that to detect credit card fraud there are many other regression is used for the prediction or forecasting the values .
approaches in Machine Learning itself. The logistic regression is mostly used for the classification
task. The J48 algorithm is used to generate a decision tree and
The research on credit card fraud detection uses both is used for the classification problem. The J48 is the extension
Machine Learning[1][2] and Deep Learning algorithms[7]. In of the ID3 (Iterative Dichotomieser). J48 is one of the most
this section, we enhance the work done in two different widely used and extensively analyzed areas in Machine
points:(i) the methods that are readily available for fraud
Learning. This algorithm mainly works on constant and
detection, and (ii) The techniques that are available to handle
categorical variables. Adaboost is one of the most widely
the imbalanced data. To handle the imbalanced data A[11]
used machine learning algorithms and is mainly developed for
some of the techniques are available. They are (a)
classification methods (b) sampling methods (c) resembling binary classification. The algorithm is mainly used to boost
techniques. Here are some of the Machine Learning the performance of the decision tree. This is also mainly used
algorithms that are used for credit fraud detection are for the classification of the regression. The Adaboost
support vector machine(SVM), decision trees, logistic algorithm is fraud cases to classify the transactions which are
regression, gradient boosting, K-nearest neighbor, etc; fraud and non-fraud. From their work they have concluded
that the highest accuracy is obtained for both the Adaboost
In 2019, Yashvi Jain, NamrataTiwari, Shripriya Dubey, and Logistic Regression. As they have the same accuracy the
Sarika jain have researched various techniques[10] for credit time factor is considered to choose the best algorithm. By
cards fraud detection such as support vector machines(SVM), considering the time factor they concluded that the Adaboost
artificial neural networks(ANN), Bayesian Networks, Hidden algorithm works well to detect credit card fraud.
Markov Model, K-Nearest Neighbours (KNN) Fuzzy Logic
system and Decision Trees. In their paper, they have
observed that the algorithms k-nearest neighbor, decision In 2019 Sahayasakila V, D.Kavya Monisha, Aishwarya,
trees, and the SVM give a medium level accuracy. The Fuzzy Sikhakolli Venkatavisalakshiswshai Yasaswi have explained
Logic and Logistic Regression give the lowest accuracy the Twain important algorithmic techniques [8] which are the
among all the other algorithms. Neural Networks, naive bayes, Whale Optimization Techniques (WOA) and SMOTE
fuzzy systems, and KNN offer a high detention rate. The (Synthetic Minority Oversampling Techniques). They mainly
LogisticRegression, SVM, decision trees offer a high detection aimed to improve the convergence speed and to solve the
rate at the medium level. There are two algorithms namely data imbalance problem. The class imbalance problem is
ANN and the Naïve Bayesian Networks which perform better overcome using the SMOTE technique and the WOA
at all parameters. These are very much expensive to train. technique. The SMOTE technique discriminates all the
There is a major drawback in all the algorithms. The drawback transactions which are synthesized are again re-sampled to
is that these algorithms don’t give the same result in all types check the data accuracy and are optimized using the WOA
of environments. They give better results with one type of technique. The algorithm also improves the convergence
datasets and poor results with another type of dataset. speed, reliability, and efficiency of the system.
Algorithms like KNN and SVM give excellent results with
small datasets and algorithms like logistic regression and In 2018 Navanushu Khare and Saad Yunus Sait have
fuzzy logic systems give good accuracy with raw and un- explained their work [5] on decision trees, random forest,
sampled data. SVM, and logistic regression. They have taken the highly
In 2019, Heta Naik, Prashasti Kanikar, has done their skewed dataset and worked on such type of dataset. The
research on various algorithms [4] like Naïve Bayes, Logistic performance evaluation is based on accuracy, sensitivity,
specificity, and precision. The results indicate that the
accuracy for the Logistic Regression is 97.7%, for Decision
Trees is 95.5%, for Random Forest is 98.6%, for SVM
classifier is 97.5%. They have concluded that the Random
Forest algorithm has the highest accuracy among the other
algorithms and is considered as the best algorithm to detect
the fraud. They also concluded that the SVM algorithm has a
data imbalance problem and does not give better results to
detect credit card fraud.
III. PROPOSED WORK
The main aim of this paper is to classify the transactions
that have both the fraud and non-fraud transactions in the
dataset using algorithms like that the Random Forest and the
Adaboost algorithms. Then these two algorithms are compared
to choose the algorithm that best detects the credit card fraud
transactions. The process flow for the credit fraud detection
problem [Figure.3.]includes the splitting of the data, model
training, model deployment, and the evaluation criteria.

978-1-7281-4876-2/20/$31.00 ©2020 IEEE 2

Authorized licensed use limited to: University of Exeter. Downloaded on June 28,2020 at 07:04:27 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Intelligent Computing and Control Systems (ICICCS 2020)
IEEE Xplore Part Number:CFP20K74-ART; ISBN: 978-1-7281-4876-2

A. Random Forest Algorithm

The Random Forest algorithm [Figure. 5]is one of the
widely used supervised learning algorithms. This can be used
for both regression and classification purposes. But, this
algorithm is mainly used for classification problems.
Generally, a forest is made up of trees and similarly, the
Random Forest algorithm creates the decision trees on the
sample data and gets the prediction from each of the sample
data. Then Random Forest algorithm is an ensemble method.
This algorithm is better than the single decision trees because
it reduces the over-fitting by averaging the result.

Figure.3 Process Flow

The detailed architecture diagram for the credit card fraud

detection system [Figure. 4.] includes many steps from
gathering dataset to deploying model and performing analysis
based on results. In this model we take the Kaggle credit card
fraud dataset and pre-processing is to be done for the dataset.
Now to prepare the model we have to split the data into the
training data and the testing data. We use the training data to
prepare the Random Forest and the Adaboost models. Then
we develop both the models. Finally, the accuracy, precision,
recall, and F1-score is calculated for bot the models. Finally
the comparison of the credit card fraud transactions more
accurately.

Figure.5 Random Forest Algorithm

Steps for Random Forest Algorithm

1. Take the Kaggle credit card fraud dataset that is

trained and randomly select some of the sample data.
2. Using the randomly created sample data now creates
the Decision Trees that are used to classify the cases
into the fraud and non-fraud cases.
3. The Decision Trees are formed by splitting the
nodes, the nodes which have the highest
Information gain make it as the root node and
classify the fraud and non-fraud cases.
Figure.4 Architecture Diagram 4. Now the majority vote is performed and the decision
Trees may result in 0 as output which includes that
these are the non-fraud cases.
5. Finally, we find the accuracy, precision, recall, and
F1 -score for both the fraud and non-fraud cases.

Random Forest algorithm

Algorithm Random Forest :

To generate c classifiers:
For i=1 to c do
Randomly select the training data D with
replacement to produce Di
Create a root node N containing Di and features in N
cell Build Tree(N) Select the features F that has the highest Information A
End for gain for further splitting
Majority Vote Gain (T,X)=Entropy (T)-Entropy(T,X) Now
to calculate the entropy we use,
Build Tree(N)
Randomly select x% of all the possible splitting
978-1-7281-4876-2/20/$31.00 ©2020 IEEE 3
Authorized licensed use limited to: University of Exeter. Downloaded on June 28,2020 at 07:04:27 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Intelligent Computing and Control Systems (ICICCS 2020)
IEEE Xplore Part Number:CFP20K74-ART; ISBN: 978-1-7281-4876-2

classifier. Adaboost algorithm can be used with short Adaboost algorithm is a powerful classifier that works well on
decision trees. The way the Adaboost is created is such that both the basic and complex problems. The disadvantage of this
initially at first the nodes are created and the tree is made, algorithm is that this algorithm is mostly sensitive to noisy
then the performance of the tree on each of the instances is data. This algorithm is also sensitive to outliers.
checked. Also, a weight is assigned. The training data that is
hard to predict is the( one ) that gives more weight. The Steps for Adaboost Algorithm
()∑
Create f child nodes 1. The Kaggle credit card fraud dataset is taken and is
For i=1 to f do trained. Randomly select some of the sample data.
Set contents f N to Di 2. Using the randomly created sample data now creates
Call Build Tree(Ni) the decision trees sequentially for classifying the
End for fraud and non-fraud cases.
3. The decision trees are formed initially. This can be
End done by splitting the node based on which has the
highest information gain, make it as the root node,
B. Adaboost Algorithm and classify the fraud and non-fraud cases.
Boosting is one of the ensemble techniques. This algorithm 4. Now calculate the error rate, performance, and
is used to build strong classifiers from weaker classifiers. This update the weights of the fraud and non-fraud
can be done by building a strong model by using a weak model transactions that are incorrectly classified.
in the series. Initially, a model is built from the training data. 5. Now majority vote is performed and the decision
Then the second model is built from the first model by trees may result as output which indicates the non-
correcting the errors that represent in the model that is created fraud cases.
before. This is a repetitive process and is continued until either 6. The decision trees may output 1 which indicates that
the maximum number of models is added or the complete it is a fraud case.
training dataset is predicted correctly. Adboost was one of the 7. Finally, we find the accuracy, precision, recall, and
most successful boosting algorithms that were developed for F1-score for both the fraud and non-fraud cases.
the binary classification.

Adaboost Algorithm

Algorithm Adaboost :
IINPUTdataset
Initialize weights, w1(n)=1/n
Create a decision tree
Select the one that has the lowest Entropy
If Incorrectly classified
Calculate Total Error (TE)= sum of up incorrectly
Classified sample weights
Calculate Performance,
For each
Incorrectly classified, increase weights:
Weights incorrect =old weight *
Correctly classified, decrease the weights:
Weight correct =old weight *
Normalized weight of each sample:
Figure.6 Adaboost Algorithm
Normalized weight =
The short name for Adaboost is adaptive boosting. It is best End for
used with weak learners. This Adaboost boosting technique End if
[Figure. 6]combines the multiple weak classifiers into a strong

978-1-7281-4876-2/20/$31.00 ©2020 IEEE 4

Authorized licensed use limited to: University of Exeter. Downloaded on June 28,2020 at 07:04:27 UTC from IEEE Xplore. Restrictions apply.
 IV. EVALUATION AND RESULT ANALYSIS

A. Dataset
The dataset, credit card fraud data is taken from the
European credit card company. The dataset is obtained from
the Kaggle. The dataset holds the transactions that are done
by the credit cardholders in the year 2013 September. The
dataset contains the transactions that are done in two days.
The data set contains 284,807 transactions in which 492
transactions are a fraud. These fraud transactions account for
only 0.172%of all the transactions. The dataset having the
input variable are converted into the numerical values by the
Figure.7 Output for Random Forest
PCA transformation. This is done due to confidentiality
reasons. The features ‘Time' and ‘Amount ‘can’t be PCA
The evaluation criteria are explained[Figure.7] and the
transformed. The class ‘Time ‘represents the difference in the
precision, recall, F1-score are the same for that of the non-
seconds elapsed between the particular transaction and the
fraud cases and differ for that of the fraud cases.
first transaction. The class ‘Amount ‘ represents the money
transaction that had occurred. Another important feature
‘Class' shows whether the transaction is fraudulent or not.
The number indication 1 shows that it is a fraud transaction
and 0 indicates the non-fraud transactions.
B. Evaluation Criteria
To compare various algorithms, we need to evaluate
metrics like accuracy, precision, recall, and F1-score. The
confusion matrix is also plotted. The confusion matrix is a 2*2
Figure.8 Confusion Matrix for Random Forest
matrix. The matrix contains four outputs which are TPR,
TNR, FPR, FNR. Measures such as sensitivity, specificity,
The confusion matrix[Figure.8] shows us that for the train
accuracy, and error-rate can be derived from the confusion
data the true positives are 190490 and false positives are 0, the
matrix. Then we that best suit to detect the credit card fraud.
true negatives are 0 and the false negatives are 330. For the
test data, the true positives are 93818 and false positives are
The output of the confusion matrix is
37, the true negatives are 7 and the false negatives are 125.
1. True Positive Rate, which can be defined as the
number of fraudulent transactions that are even
classified by the system as fraudulent.
2. True Negative Rate, which can be defined as the
number of legitimate transactions that are even
classified as legitimate by the system.
3. False Positive Rate, which can be defined as a
number of the legal transactions which are wrongly
classified as fraud.
4. False Negative Rate is defined as the transactions
that are fraud but are wrongly classified as legal.

The Receiver Operating Characteristics curve is created by

plotting the TPR against the FPR. This can be done at various Figure.9 ROC curve for Random Forest
thresholds. ROC curve is a graph in which the FPR is the
horizontal axis and the TPR is the vertical axis. The graph Now the dataset is applied for the Adaboost algorithm. The
under the ROC curve is the AUC. results are obtained similar to that of the Random Forest
Algorithm.
C. Results Analysis
The confusion matrix and the ROC curve is plotted for both
the algorithms. The dataset, when applied for different
algorithms, gives different outputs. Firstly we apply the dataset
for the random forest model and the results are as below:
Figure.10 Output for Adaboost
The evaluation criteria[Figure.11] shows us that the
evaluation criteria like the precision, recall, and F1-score differ
less in the case of the non-fraud cases and differ greatly in
those of the fraud cases.

Figure.11 Confusion Matrix for Adaboost

 The confusion matrix [Figure.11] shows us that for the
V. CONCLUSION
train data the true positives are 190464 and false positives are
120, the true negatives are 26 and false negatives are 201. For Even though there are many fraud detection techniques we
the test data, the true positives are 93811 and false positives can’t say that this particular algorithm detects the fraud
are 65, the true negatives are 14 and false negatives are 97. completely. From our analysis, we can conclude that the
accuracy is the same for both the Random Forest and the
Adaboost algorithms. When we consider the precision, recall,
and the F1-score the Random Forest algorithm has the highest
value than the Adaboost algorithm. Hence we conclude that
the Random Forest Algorithm works best than the Adaboost
algorithm to detect credit card fraud.

VI. FUTURE SCOPE

From the above analysis, it is clear that many machine learning
algorithms are used to detect the fraud but we can observe that
the results are not satisfactory. So, we would like to implement
Figure.12 ROC curve for Adaboost deep learning algorithms to detect credit card fraud accurately.

Now the comparison of the random forest and the Adaboost REFERENCES
algorithms is shown [Figure.12]. The two algorithms have the 1. Adi Saputra1, Suharjito2L: Fraud Detection using Machine
same accuracy but the precision, recall, and the F1-score of the Learning in e-Commerce, (IJACSA) International Journal of
Advanced Computer Science and Applications, Vol. 10, No. 9,
two algorithms differ. The random forest algorithms have the 2019.
highest precision, recall, and F1-score. 2. Dart Consulting,Growth Of Internet Users In India And Impact
On Country’s Economy: https://www.dartconsulting.co.in/market -
news/growth-of-internet-users-in-india-and-impact -on-countrys-
economy/
3. Ganga Rama Koteswara Rao and R.Satya Prasad, “ - Shielding
The Networks Depending On Linux Servers Against Arp
Spoofing, International Journal of Engineering and
Technology(UAE),Vol. 7, PP.75-79, May 2018, ISSN No:
2227-524X, DOI - 10.14419/ijet.v7i2.32.13531.
4. Heta Naik , Prashasti Kanikar: Credit card Fraud Detection
based on Machine Learning Algorithms,International Journal
of Computer Applications (0975 – 8887) Volume 182 – No.
44, March 2019.
5. Navanshu Khare ,Saad Yunus Sait: Credit Card Fraud
Detection Using Machine Learning Models and Collating
Machine Learning Models, International Journal of Pure and
Applied Mathematics Volume 118 No. 20 2018, 825 -838
ISSN: 1314-3395.
6. Randula Koralage, , Faculty of Information Technology,
Figure.13 Comparision of Algorithms University of Moratuwa,Data Mining Techniques for Credit
Card Fraud Detection.
7. Roy, Abhimanyu, et al:Deep learning detecting fraud in credit
card transactions, 2018 Systems and Information Engineering
Design Symposium (SIEDS), IEEE, 2018.
8. Sahayasakila.V, D. Kavya Monisha, Aishwarya, Sikhakolli
VenkatavisalakshiseshsaiYasaswi: Credit Card Fraud Detection
System using Smote Technique and Whale Optimization
Algorithm,International Journal of Engineering and Advanced
Technology (IJEAT) ISSN: 2249-8958, Volume-8 Issue-5,
June 2019.
9. Statista.com. retail e-commerce revenue forecast from 2017 to
2023 (in billion U.S. dollars). Retrieved April 2020, from India
: https://www.statista.com/statistics/280925/e-commerce-
revenueforecast-in-india/
10. Yashvi Jain, NamrataTiwari, ShripriyaDubey,Sarika Jain:A
Comparative Analysis of Various Credit Card Fraud Detection
Techniques, International Journal of Recent Technology and
Engineering (IJRTE) ISSN: 2277-3878, Volume-7 Issue-5S2,
January 2019.

11. Yong Fang1, Yunyun Zhang2 and Cheng Huang1, Credit Card
Fraud Detection Based on Machine Learning, Computers,
 Materials & Continua CMC, vol.61, no.1, pp.185
-195, 2019.
12. Kaithekuzhical Leena Kurien, Dr. Ajeet Chikkamannur:
Detection And Prediction Of Credit Card Fraud Transactions
Using Machine Learning , International Journal Of
Engineering Sciences & Research Technolog.