International Journal of Computer Trends and Technology Volume 68 Issue 6, 77-81, May 2020

ISSN: 2231-2803 / https://doi.org/10.14445/22312803/IJCTT-V68I6P113 © 2020 Seventh Sense Research Group®

Review Article

Review on Credit Card Fraud Detection Using

Machine Learning Algorithms
Pooja1, Ashlesha2
1
Research Scholar (Network engineering), J.C. Bose University of Science and Technology
YMCA at Faridabad, Haryana - India
2
Assistant Professor(Computer engineering department), J.C Bose University of Science and Technology, YMCA at
Faridabad – Haryana – India.

Received Date: 06 May 2020

Revised Date: 20 June 2020
Accepted Date: 23 June 2020

Abstract -- The advancement of new technologies and the transactions, there is also very growth in fraud activities. In
fast-growing of technological development have generated a given transaction, credit card fraud is a term used for
new possibilities as well as imposed new challenges. theft and fraud observed by using credit cards as a source
Fraud, the biggest challenge for business and of fraud funds. Regular use of credit card transactions for
organization, emerge with new technologies to take new obtaining Goods and Services assisting online or card
and distinctive forms that are hidden and tougher to swiping method leading to continuous growth in the online
identify than the conventional forms of this crime. Credit transaction by use of credit and debit card increasing to the
card frauds also grow up along with growth in technology. world of relaxing expenses. These frauds of credit cards
It also noticed that financial fraud is extremely growing in caused greater damage to the customers and service
the global communication improvement. It is being providers. This is becoming very bad in the coming future.
admitted every year that the loss because of these types of This person finds and adapt to changes in technology and
fraudulent activities is billions of dollars. These activities find new and easy clever ways to do these fraudulent
are performed so gracefully that they look similar to activities. Fraud due to these activities are very dangerous
original transactions. Simply using of pattern matching and risky. The very smart fraudster creates many identities
technique and simple method is really not useful for and does online transactions without being caught. It is
detecting these fraudulent activities. A well planned and very hard to detect these fraud activities as these activities
systematic method has become the need for all businesses look real, and datasets are not available easily. The bank's
and organizations to minimize chaos and carry it out in owner and service provider are not interested in sharing the
place. Several techniques have been evolved based on dataset for experiments.
Artificial intelligence, Machine learning, Data mining, Fraud can be interpreted as fraud committed to
Genetic programming, Fuzzy logic etc.. for detecting presenting the financial statements of a company. Fraud
credit card fraudulent activities. Besides this technique, (fraud) is intentional fraud that gives a profit to the
the K-Nearest Neighbour algorithm and outlier detection company and causes losses to the company. Fraud
methods are implemented to optimize the best solution for committed by several parties is one of the most interesting.
the fraud detection problem. These techniques proved to In general, there are three things that encourage the
minimize the false alarm rates and increase the fraud transfer of fraud (pressure), opportunities (opportunities)
detection rate. and justification for the actions taken (rationalization).
Encouragement (pressure) is an impetus arising from a
Keywords – Credit Card, Fraud Detection, Machine desire to get a better life. Luxury supported by poor
Learning, Logistic Regression, K-Nearest Neighbour. economic conditions also encourages the environment for
a luxurious lifestyle. Opportunity (opportunity) is a
I. INTRODUCTION deception factor used by his weak party Rationalization
Credit card fraud detection is a way of recording the (rationalization) is a reason made by the party who
selling and buying behaviour of the customer during a committed fraud.
certain period. A credit card is just a plastic card given to
customers As a payment system. Cardholders can buy The types of fraud can be grouped into 3, namely:
Goods and Services on the basis of a promise to pay for  Employee fraud (employee fraud), namely fraud
using these Goods and Services. Security of credit cards committed by employees in a work organization.
depends upon the physical security and on the privacy of  Management fraud is fraud committed by
credit card Numbers. Globalization And due to increasing management using financial statements or financial
use of the internet for Online Shopping has to Result in a transactions as fraud.
substantial increase in credit card transactions around the
world. Due to the growth in the use of credit card

This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)

 Pooja & Ashlesha / IJCTT, 68(6),77 - 81, 2020

 Computer Fraud (Computer Fraud) is fraud input layer, hidden layer and output layer. Each node uses
committed to recording computers that contain an activation function. The activation function calculates
operational or bookkeeping records in a company. the weighted sum of its inputs and adds bias. This allows
us to decide which neuron should be removed and not
A. Types of credit card fraud considered in outside connections.
a) One is application fraud, where an individual will KNN is the K-nearest neighbour algorithm that is used
forge an application in order to receive a credit card. largely in fraud detection systems using supervised
They give incorrect information about /her financial learning techniques. In KNN, a newly arrived data
status and receive a credit card. classified depend upon the knn category. This method was
b) Second is assumed identity, where an individual first utilized by Aha, Albert and Kibler in the year 1991.
assumes someone identity and forge a name with a The outcome of KNN is based upon three factors:
temporary address.  The distance metric is used to decide the nearest
c) The third is financial fraud, where individuals apply neighbours.
for a credit card with her/ own name, but  The distance rule that is used for the classification
information related to them is false. This happens from K- nearest neighbour.
when the individual wants more than the credit that  The number of neighbours considered to classify
he currently has. the new sample.
d) The fourth type is skimming technology, where only
the purpose is to collect and store information on C4.5 algorithm is one of the classification algorithms
the credit card. Magnetic card skimming is a small introduced by J. Ross Quinlan (1996) as an improved
handheld device used for this purpose. version of ID3. C4.5 algorithm is the development of the
e) The last type is never getting an issue where Decision Tree classification tree or decision tree.
individuals steal the card while the card is in use. Primarily, the selection of the breaking point (node) in
These types use holder mail for stealing the card. this algorithm is based on the Gain calculation to induce
the tree to be formed. In the classification process of the
B. Data mining Fraud Detection Technique C4.5 algorithm, there are destination variables that are
a) Supervised Learning for Fraud Detection usually grouped with certainty. Next, a decision tree
In this method, all recorded datasets are classified model will be formed by calculating the probability of
into fraudulent and non-fraudulent earlier. Machine each record for each feature. Every data to be tested using
classify records in accordance with training given. This the C4.5 algorithm needs to be divided into training data
method only identified the fraud that has already and testing data for each feature and label.
happened, and the system trained already about these.
Few of supervised techniques used are b) Unsupervised Learning for Fraud Detection
Logistic regression is one of the most popular This technique only identify the likelihood of a few
classification algorithms in machine learning. The records being more fraudulent than other records without
logistic regression model describes the relationship any assurance. There is no classification earlier. All
between predictors that can be continuous, binary, and records are separately handled.
categorical. The dependent variable can be binary. The KMeans algorithm is the best-known partitioned
Based on some predictors, we predict whether classification algorithm which is a simple method for
something will happen or not. We estimate the estimating the average (vector) of a K groupset. Kmeans
probability of belonging to each category for a given set is the most widely used among all clustering algorithms
of predictors. because of its efficiency and simplicity. KMeans is a
Naive Bayes is one of the supervised learning well-known and widely used grouping algorithm.
algorithms in which there are no dependencies between KMeans is one of the simplest grouping algorithms in
attributes. It's based on the Bayes theorem. Depending machine learning that can be used to automatically
on the type of distribution, there are the following recognize groups of similar objects in training data.
algorithms: Gaussian distribution, Multinomial
distribution, Bernoulli distribution. In this research, II. LITERATURE REVIEW
Bernoulli distribution is used for detecting fraud N. Malini, Dr M. Pushpa, the author of a research
transactions. paper[1], says that K-Nearest Neighbors (KNN) and
Random forest is an algorithm that can be used in outlier detection techniques are very efficient in fraud
both classification and regression problems. It consists detection. This technique has proven useful in minimizing
of many decision trees. This algorithm gives better false alarm rates and increasing fraud detection rates. The
results when there is a higher number of trees in the author takes a new object set and firstly take one nearest
forest and prevents the model from overfitting. Each neighbour and then increase the number of nearest
decision tree in the forest gives some results. These neighbour one each time. The author takes the simple
results are merged together in order to get the more example of a positive and negative point. By taking one
accurate and stable prediction. neighbour vote gone to positive sign and by taking two
Multilayer perceptron is feeding forward artificial neighbours, both positive and negative points have equal
neural network that consists of minimum 3 layers of nodes: voting. By taking 5 neighbours, the author finds that voting

78
 Pooja & Ashlesha / IJCTT, 68(6),77 - 81, 2020

for the positive sign is more than a negative sign. So method. These methods are useful in detecting fraudulent
voting goes to the positive class, and the new object goes activities.
to the positive class. The author suggests taking a value of F. Ghobadi, M.Rohani the author of the research
k always an odd number. paper[6], make a comparison of neural network and
J.O. Awoyemi et al., the authors of the research paper various ensembles techniques and finds neural networks
[2], finds KNN algorithm performed well where the are better performance in comparison of ensembles
authors tested and compared it with other classical techniques. Three classifier models based on Neural
algorithms used for credit card fraud detection. The author network, k-nearest neighbour and logistic regression are
does the comparative performance of Naïv Bayes, K developed. To evaluate these models, 70% of the dataset is
Nearest neighbour and Logistic regression models in the used for training, while 30% is set aside for validating and
binary classification of imbalanced credit card fraud data. testing. Accuracy, sensitivity, specificity, precision,
Three classifiers based on different machine learning Matthews correlation coefficient (MCC) and balanced
techniques (Naïve Bayes, K-nearest neighbours and classification rate are used to evaluate the performance of
Logistic Regression) are trained on the real-life of credit the three classifiers. The accuracy of the classifiers for the
card transactions data and their performances on credit original 0.172:99.828 dataset distribution, the sampled
card fraud detection evaluated and compared based on 10:90 and 34:66 distributions are presented in Tables 1, 2
several relevant metrics. The performances of the three and 3, respectively. An observation of the metric tables
classifiers are examined on the two sets of data shows that there is a significant improvement from
distributions using accuracy, sensitivity, specificity, the sampled dataset distribution of 10:90 to 34:66 for
precision, balanced classification rate and Matthews accuracy, sensitivity, specificity, Matthews correlation
Correlation coefficient metrics. The dataset contains coefficient and balance classification rate of the classifiers.
284,807 transactions, where 492 transactions were frauds, This shows that a hybrid sampling(under-sampling and
and the rest were genuine. Considering the numbers, we over-sampling) on a highly imbalanced dataset greatly
can see that this dataset is highly imbalanced, where only improves the performance of binary classification. The
0.173% of transactions are labelled as frauds. true positive, true negative, false positive and false
Z. Kazemi, H. Zarrabithe author of the research negative rates of the classifiers in each set of unsampled
paper[3], made a comparison of various deep learning and sampled data distribution is shown in Tables 4 5and6.
techniques. The author compares certain machine learning Logistic regression is the only technique that did not show
algorithms for the detection of fraudulent transactions. better improvement in false-negative rates from the 10:90
Hence comparison was made, and it was found that the to 34:66 data distribution. However, it showed the overall
Random Forest algorithm gives the best results, i.e. best best performance in the un-sampled distribution.
classifies whether transactions are fraud or not. This was Venkata Ratnam Ganji et al. [7]the author use concept
established using different metrics, such as recall, accuracy, of data stream outlier detection algorithm, which is based
and precision. on anti k nearest neighbour for credit cards fraud detection,
N. Kalaiselvi, S. Rajalakshmi, J. Padmavathi Authors whereas traditional methods need to scan the dataset many
of paper[4] used neural networks in order to demonstrate times to find fraudulent transactions, which is not suitable
improvement in results when ensemble techniques are used. for data stream surroundings. This method makes it easier
For analysis, the author used a sample set of 5,850 fraud to stop fraudulent transaction that happens lost and stolen
transactions and 542,858 legal transactions, ordered by credit card .validation check and detects errors in a
their timestamps. It should be noted that the mining sequence of numbers, which also helps to detect valid and
algorithm has a high runtime complexity. So author used invalid numbers easily.
only 30,000 of the legal transactions. The resulting values Abhinav Srivastava, the author of the research
for the confidence were compared to the whole set of paper[8], uses the ranges of the transaction amount as an
transactions. The author uses only fraud occurrences attribute in the HMM. The author has suggested a method
of .1%. The simple constant diagnosis "transactions are no for finding the spending profile of cardholders. It is also
fraud" will have a success rate of99.9%. To compete with discussed how the HMM can identify fraudulent
this trivial diagnosis, the task of diagnosing a transaction is transactions. The simulation results show the advantages
not easy-to-do if the author use only the analogue data; all of using HMM, and learning the profile of the cardholder
transactions patterns characterized by n symbolic and m plays an important role in analyzing fraudulent cases. The
nalog features are projected from the n+m dimensional Result also shows that 80% of the results are accurate, and
space into the m dimensional space. Generally, this results the system is scalable for large data set as well.
in overlapping classes and, therefore, in diagnostic success Divya. Iyer et al. [9] the author uses Hidden Markov
far worse than 99.9%. Model (HMM) to detect credit card transaction frauds. The
Mrs C. Navamani, M. Phil, S.Krishnan author of the training set is tuned with the normal behaviour of the
research paper [5], find outlier detection techniques are cardholder. So if a credit card transaction is rejected by the
very efficient in fraud detection. This technique has proven trained HMM, then that transaction is said to be fraudulent.
useful in minimizing false alarm rates and increasing fraud Care is to be taken that valid and genuine transactions are
detection rates. The author studies different outlier not considered fraud. The author also compares various
detection techniques. Apply this technique for fraud methods with the proposed methods to prove that HMM is
detection. Credit card fraud can be solved using this outlier much preferred than the other methods.

79
 Pooja & Ashlesha / IJCTT, 68(6),77 - 81, 2020

K.RamaKalyani et al. [10] create test data through parameters. In addition, MO-GAAL can easily handle
which fraudulent activities are detected. This algorithm is various types of clusters and high irrelevant variables.
also called an optimization technique based on genetic and K.T.Divya, N.Senthil Kumaran [15] depicted that,
natural selection in great computational problems. The The outlier detection approach is based on distance
author proposes a method to detect credit card fraud, and learning for category attributes (DILCA), a distance
the results are validated using the principles of this learning framework is introduced. The key intuition of
algorithm. The purpose of detecting fraud cases is to DILCA is that the distance between two categorical
declare it to the client and the service provider. attribute values can be determined in a way where they
occur together with other attribute values in the data set.
Renu et al. [11] proposed a fraud detection method The classic KNN produces superior data utility but raises a
that involves monitoring the activities of populations to higher computational overhead. In addition, dimension
observe and predict undesirable behaviour. Undesirable reduction techniques used in the occupational health
behaviour is a set of several habits like intrusion, fraud, dataset are used.
delinquency and default.
Fraud
K.Swapna, Prof. M.S. Prasad Babu [12] depicted that, Detection Advantage Disadvantage
Design a liver diagnosis system automatically to detect Technique
early and accurately to reduce deaths caused by liver Work well with
disease and analyze data sets to understand the system to Can not handle non-
linear data and
design a liver diagnosis system automatically to detect linear data
detect fraud by
early and accurately to reduce deaths caused by liver Logistic
creating a
disease and analyze data set to understand the use system. Regression can not detect fraud
probabilistic
at the time of
formula for
Heta Naik [13] depicted that, The increase in online transaction
classification
transactions is directly proportional to the increasing
number of frauds. In this paper, various algorithms such as Algo very complex
K-Nearest Neighbor, Random Tree, AdaBoost and Can apply on
even a small change
Logistic Regression are some challenges which include Decision Tree linear as well as
can change the whole
distinguishing between normal transactions and fraud that on non-linear data
tree structure
seem very similar to each other. The parameters to detect
these transactions are Time, Number and Frequency of
It can detect
Transactions. In this paper, four different KNN algorithms,
fraudulent
AdaBoost, Random tree and Logistic Regression, are
activities at the It can not detect
compared for fraud detection mechanisms. Logistic Hidden
time of fraud at some initial
regression is better compared to other algorithms. This Markov Model
transactions and transactions
model is used for unbalanced credit card fraud data. All of
reduce the false-
these algorithms do not apply to fraud detection at the time
positive ratio
of the transaction.
The number of the
Yezheng Liu et al. [14] depicted that, Approach outlier parameter is to be set
detection as a binary classification problem by taking before training
potential outliers from a uniform reference distribution. started
However, due to the scarcity of data in high-dimensional
space, a number of potential outliers may fail to provide Can detect fraud no certain method till
information to help the classifier draw a line that can Artificial activity at the now for deciding
effectively separate outliers from normal data. To neural network time of online optimal topology for
overcome this, propose a Single Objective Objective transaction a particular problem
Active Target (SO-GAAL) Active Learning method for
network working
outlier detection, which can directly generate potential
depends upon the
informative outliers. Proposed a new SOGAAL outlier
interconnection of
detection algorithm, which can directly produce potentially
neurons
informative outliers, to overcome the lack of information
caused by curse dimensions. Extending the GAAL Can not detect fraud
structure of a single generator (SO-GAAL) to several No requirement of at the time of the
generators with different purposes (MO-GAAL) to prevent K-nearest predictive model transaction, and
generators from falling into a collapsing problem mode. neighbour before accuracy depends on
Compared to some sophisticated outlier detection methods, classification the distance
MO-GAAL achieves the best average rating in real-world measured
datasets and shows strong robustness for various

80
 Pooja & Ashlesha / IJCTT, 68(6),77 - 81, 2020

III. CONCLUSION Conference on Computation of Power, Energy, Information and

Communication (ICCPEIC) International conference on the
Credit card fraud has become very large these days. computation of power, energy, Information and Communication
To progress safety measures of the monetary transaction (ICCPEIC) (2018) 191-196.
systems in a habitual and effectual way, structuring a [5] Mrs C. Navamani, M. Phil, S. Krishnan, Credit Card Nearest
Neighbor Based Outlier Detection Techniques, International Journal
precise and well-organized credit card scam detection of Computer Techniques 5 (2) (2018)
system is one of the essential functions for money [6] F. Ghobadi, M. Rohani, Cost Sensitive Modeling of Credit Card
transactions. By performing oversampling and extracting Fraud using Neural Network strategy, Signal Processing and
the principal direction of the data, we can use our KNN Intelligent Systems (ICSPIS), International Conference of
(2016) 1-5.
method to determine the anomaly of the target instance. [7] Venkaata Ratnam Ganji, credit card fraud detection using the anti-K
Hence the KNN method can suit for detecting fraud with nearest neighbour algorithm, International Journal on Computer
the limitation of memory. In the meantime, the outlier Science and Engineering(IJCSE) 4 (6) (2012) 1035-1039
detection mechanism helps to detect credit card fraud [8] Abhinav Srivastava, Amlan Kundu, Shamik Sural, and Arun K.
Majumdar Credit Card Fraud Detection Using Hidden Markov
using less memory and computation requirements. Model 5(1) (2008).
Especially outlier detection works fast and well on large [9] Divya.Iyer, Arti Mohanpurkar, Sneha Janardhan, Dhanashree
online datasets. But compared with power methods and Rathod, Amruta Sardeshmukh Credit card Fraud Detection using
other known anomaly detection methods, experimental Hidden Markov model 978-14673- 01268/11/$26.00_c 2011 IEEE
[10] K.RamaKalyani, D.UmaDevi Fraud Detection of Credit Card
results prove that the KNN method is accurate and Payment System by Genetic Algorithm 3(7) (2012)
efficient [11] Renu, Suman. Analysis on Credit Card Fraud Detection Methods.
International Journal of Computer Trends and Technology (IJCTT)
REFERENCES 8(1) (2014) 45-51. www.ijcttjournal.org. Published by Seventh
[1] N. Malini, Dr M. Pushpa, Analysis on Credit Card Fraud Sense Research Group.
Identification Techniques based on KNN and Outlier Detection , [12] Swapna K, and Babu M S PInternational Journal of Electrical &
Advances in Electrical, Electronics, Information, Communication Computer Sciences IJECS-IJENS A Framework for Outlier
and BioInformatics(AEEICB), Third International Conference. Detection Using Improved Bisecting KMeans Clustering
(2017) 255-258. Algorithm.(2017).
[2] J. O. Awoyemi, A. O. Adentumbi, S. A. Oluwadare, Credit card [13] Naik H International Journal for Research in Applied Science &
fraud detection using Machine Learning Techniques: A Engineering Technology (IJRASET) Credit Card Fraud Detection
Comparative Analysis, Computing Networking and Informatics for Online Banking Transactions (2018).
(ICCNI), International Conference on (2017) 1-9. [14] Yezheng Liu, Zhe Li, Chong Zhou, Yuanchun Jiang, Jianshan Sun,
[3] Z. Kazemi, H. Zarrabi, Using Deep Networks for Fraud Detection in Meng Wang and Xiangnan He / Generative Adversarial Active
the Credit card Transactions, Knowledge-Based Engineering and Learning for Unsupervised Outlier Detection (2019)
Innovation (KBEI), IEEE 4th International Conference on (2017) [15] Divya K T, Kumaran N S International Research Journal of
630-633. Engineering and Technology (IRJET) Improved Outlier Detection
[4] N. Kalaiselvi, S. Rajalakshmi, J. Padmavathi, Credit card fraud Using Classic Knn Algorithm (2016).
detection using learning to rank approach, Internat International

81