Anomaly Detection in Credit Card

Transactions Using Machine Learning

Techniques

By

1
Contents
1. Introduction..................................................................................................................................3

1.1 Background to the Problem....................................................................................................3

1.2 Evidence to Support the Problem...........................................................................................4

1.3 Aim and Objectives................................................................................................................4

1.4 Significance............................................................................................................................5

1.5 Scope of the Study.................................................................................................................5

1.6 Limitation...............................................................................................................................6

2. Literature Review........................................................................................................................7

2.1 Definitions..............................................................................................................................7

2.2 Theories Associated to the Research Objectives....................................................................8

2.3 Empirical Findings.................................................................................................................9

2.4 Critical Review......................................................................................................................9

3. Methodology..............................................................................................................................11

3.1 Conceptual Framework........................................................................................................11

3.2 Hypothesis/Propositions.......................................................................................................11

3.3 Operationalization of the Variables......................................................................................11

3.4 Research Design...................................................................................................................11

3.5 Methods of Sampling, Data Collection, and Analysis.........................................................12

3.6 Proposed Models/Analysis Methods....................................................................................12

3.7 Ethical Considerations.........................................................................................................12

3.8 Time Plan.............................................................................................................................12

4. Conclusion.................................................................................................................................13

5. References..................................................................................................................................14

2
1. Introduction

1.1 Background to the Problem

Credit card fraud has been one modern danger in the recent pasts with the increase in the credit
card usage with millions of transactions. By the very nature of the tremendous amount of
transactions and given the fact that fraudulent activities are often intricate in their nature, it is
difficult for the more basic form of manual inspection to prevent the occurrence of fraudulent
transactions. This has placed a huge economic strain on both the business and consumer front
and eroded the very fabric of trust in financial institutions.

The number of transactions is another concern because the tens of thousands of transactions
taking place every day render the possibility of finding meaningful use from these transactions
impossible for human-specialist. Furthermore, the dataset used in this project underscores the
challenge with a high imbalance: Even in the worst-case scenario of 284,807 transactions, only
492 are fraudulent in our case, implying an odds ratio of 0. 17% fraudulent transactions. This
situation makes it difficult to call for a simple or easily understandable detection. There are high
costs of finance owing credit card fraud incidences, where the Federal Trade Commission (FTC)
recorded 114,348 for the year 2023. (rajeevvhanhuve, n. d. )

Earlier this method was applied making observations directly but it failed because of increased
numbers of transactions. They are inflexible as they are designated by specific rules and patterns
and stand helpless when facing new types of frauds. This goes a long way in showing the
necessity of developing high-end technology, which can monitor for any fraudulent transactions,
immediately.

Discrimination has been identified as a major drawback of self-learning, and machine learning
provides a powerful antidote to this problem. Machine learning techniques also help to sort big
financial data and recognize fraudulent activity and genuine transactions besides profound
knowledge and experience. This is especially useful in averting fraud cases which are usually
determined as unusual to normal business operations.

Nevertheless, integrating it into the system to assist in antigraffiti efforts is not without its
difficulties. The distinguishing factor here is that the high imbalance of data in the case of
malicious webservers is a severe threat to good performance and results in machine learning
3
models. One of the challenges facing the model is that scammers learn how to avoid detection by
the model making it crucial to incorporate better models that are capable of detecting anomalous
lograms a. Moreover, checking for fraudulent transactions in real-time would act as a measure of
containing any losses that may have resulted from the fraudulent activities that affect the
customers. In conventional machine learning, the models could be heavy and slow in order to
perform the transaction in real time. (jen, 2024)

Credit card fraud detection is a challenging task due to the high level of sophistication involved
in implementing technological interventions to prevent fraud. The use of machine learning in this
case presents a promising solution to the problem of identifying fraudulent transactions, that is, it
can become a powerful tool in this regard. Requirements such as having sound and flexible
techniques, the ability to detect credit card fraud in real-time, and handling cases of skewed data
are important and hence are primary when it comes to credit card fraud detection techniques.

1.2 Evidence to Support the Problem

The following are some pieces of evidence appreciate the extent of credit card fraud and the
challenges inherent in its detection: First, an essential issue for this project is the lack of data
bias: there are substantially more minorities than whites in this dataset. A total of 284,807 initial
transactions are processed; out of them only 492 are verified as fraudulent one, thus 284806:492
= 582:1 or strictly 1 for 582. 17% fraudulent transactions. This is quite an irrational distribution
and with this sort of discrepancy it becomes hard to teach the detection models not to be biased.
(Leleko, 2024)

Secondly, it can be also stated that the amount of data introduced to this process is extremely
large to process it during one day. Transaction data is gargantuan; it can be daunting for human
specialists to decipher what looks suspicious and what does not from the large transaction data
sets. Due to this high volume it requires the use of sophisticated technologies to capture, store
process as well as analyze the data.

Finally, early marking techniques of these scams change over time with increased efforts to
develop models to detect them. The fact that fraudsters are also changing their tactics to evade
detection shows the importance of coming up with effective and more importantly versatile
means of identifying anomalies. Static techniques which are conventional in detecting credit card

4
are unable to provide the services that are required by these modern techniques; therefore, the
world needs more dynamic and efficient methods to fight against credit card fraud.

1.3 Aim and Objectives

 Aim

Develop a machine learning model that can accurately detect anomalous transactions in
credit card data.

 Objectives

 Identify the most effective machine learning algorithms for anomaly detection in
credit card transactions.

 Evaluate the performance of these algorithms in detecting fraudulent transactions.

 Develop a robust and adaptive model that can handle the challenges posed by
imbalanced data and adaptive techniques used by scammers.

1.4 Significance
Therefore, it is ridden with implications on why proper credit card fraud detection system should
be developed. First, it highlights that accurate identification of fraudulent transactions can
enormously minimize credit card businesses and customer’s loss. Thus having realized and
quickly prevented fraudulent activities, significant amounts of money will be saved. (Alvarez,
2019)

Second of all, the possible increased identification rate shows that more security assets for the
consumers can be offered by developing a better model. This boost in security helps to minimize
the instances of fraud in transactions and helps to create more confidence among the consumers
when dealing with the financial markets.

1.5 Scope of the Study

The focus of this study will be to implement training and test datasets and test them through
building and improving the machine learning models for the detection of credit card transactions
anomalies. The first component of the study entails data acquisition that scouts and preprocesses
5
the data set referring to the credit card transactions. This step is very important to prevent data
that is irrelevant or of poor quality, that will make the task of training the machine learning
model difficult.

The next step is model development; many details of the Machine Learning models will be
established and employed solely for the purpose of anomaly detection. This entails identifying
proper algorithms that can be applied as well as adjusting the constants of the applied algorithms
in order to efficiently identify fraudulent transactions (Casler, 2020).

After the model development process is complete, this study will then move on to assess the
performance of the models that have been integrated in the study. The evaluation will be made
based on different indicators such as precision, recall, and F1-scores through which the degree of
accuracy with which different models is capable of detecting fraudulent transactions will be
determined. [42] Such statistics offer an evaluation of the efficiency of the models in the area of
anomaly recognition.

1.6 Limitation
 The availability of high-quality, well-labeled data for credit card transactions is limited.

 The complexity of the models may lead to overfitting, which can negatively impact their
performance.

 Scammers may adapt to the detection models, making it essential to continuously update
and refine the models.

6
2. Literature Review

2.1 Definitions
Credit card fraud is one of the major concerns in the finance industry that relates to the
unauthorized use of a credit card and an action with value. …that involve different types of
fraud: Identity theft, card skimming, and phishing. The mention of a credit card fraud in this
research work is synonymous to a payment card fraud defined as any unlawful action with a
payment card including a credit card or debit card. (Sulaiman, 2022)

Credit card fraud impersonation is one of the popular credit card scams where an identity theft is
conducted using an individual’s information to create new credit accounts or initiate
unauthorized transactions in the vicitm’s name. This can be achieved through phishing,
skimming and a user contributing to the transfer of information often a friendly user.

Skimming is a type of identity theft where a small device is placed over the credit card slot in an
ATM or any point-of-sale devices, to read and record the information on the track. It is then used
to clone cards or perform other illicit transactions Indeed the following of such illicit activities is
usually followed by advance planning and execution of a systematic approach. (jen, 2024)

Another form often employed by fraudsters is phishing whereby they send fake emails with links
to look like genuine websites where they believe people will entering their credit card details.
This means that somebody sends fake e-mail, site, or telephone call to lure people to reveal their
personal details including credit card numbers or code.

Card cloning entails the duplication of the information for the stolen credits cards using
skimming tools. These cloned cards are then utilized to perform other unlawful charges.

Credit card details are stolen by cybercriminals by hacking services, or even by putting in a virus
on the computer of the individual or business in question. This may happen through e-mails
which one is likely to click on a link that contains a virus, having downloaded a software with a
virus or even connecting to a wrong network.

7
Social engineering is a technical deception commonly employed by scammers to influence
people and cause them to reveal their credit card details or other personal information. This can
be in form of phone calls, emails or face to face contact by the use of interpersonal
communications.

Fraud with credit cards is another fraudulent problem that encompasses different kinds of
fraudulent situations. To know how to effectively combat credit card fraud, it is pivotal to grasp
these definitions and techniques relied on by these fraudsters.

2.2 Theories Associated to the Research Objectives

The research objectives of this study are closely aligned with two prominent theories in the field
of behavioral and technology adoption research: The theory of choice for this study is the Theory
of Planned Behavior (TPB) and the other relevant theory is the Technology Acceptance Model
(TAM). (Alvarez, 2019)

The Theory of Planned Behavior (TPB) is one of the common applied models which provides
the exposure of three variables including attitude, subjective norm, and perceived behavioral
control in relation to an individual’s behavioral intention. (rajeevvhanhuve, n. d. ) Thus, in the
case of credit card fraud, TPB perspective would offer a very interesting angle in terms of
understanding the factors that make people engage in fraudulent behaviours.

TPB posits that intentions of credit card fraud, perceived behavioral control or perceived
availability of opportunities to engage in the undesirable behavior, subjective norm or perceived
social pressures not to engage in credit card fraud and personal attitude towards the behavior, all
predict the intention of the person to engage in credit card fraud. For instance, when an
individual have a positive attitude towards credit card fraud, with the perception that others are
also engaged in the act and have the capability and opportunity to perform the act, the actual
intention to engage in credit card fraud will be strong. (Leleko, 2024)

Another theory that is also related in this study is the Technology Acceptance Model (TAM) that
posits that a specific stock subjective attitude on perceived ease of use and perceived usefulness
of a certain system affects the system usage. TAM can be useful in the context of credit card
fraud detection systems in enhancing understanding of the factors that define the perceived ease

8
of use and perceived usefulness, and how these define the ability of organisations to adopt these
systems. (jen, 2024)

Acceptance by credit card holders depends on usage of the system as gauged by ease of use and
perceived effectiveness thus implying that if the holders of credit card feel that the fraud
detection system is effective in weeding out fraudulent activities then the system is likely to
success in preventing credit card fraud.

2.3 Empirical Findings

The current trend of credit card fraud detection has already tried a number of ML and DL
techniques. Some of these studies have used performance measures that varied a lot, pointing out
that these approaches are efficient in identifying fraudulent transactions.

For example, a paper that sorted data using XGBoost, a gradient boosting decision tree model,
had a 95 percent accuracy. 64% with precision as 0.82 and recall as 0.67 and finally F1-score as
0.7149. 93 in identifying Credit-card fraud. The precision which is one of the classifiers of the
F1- score measures the accuracy of the model when identifying the fraudulent transactions and
the recall measures how many of the wrongly identified fraudulent transactions were actually
accurate the two measures show that the model consistently identified almost all the fraudulent
transactions while at the same time minimizing on wrongly identifying normal transactions as
fraudulent. (Laplante, 2014)

Another study involved the use of neural network analysis and the confused rate that was
recorded was 93. 5% which corresponds to the accuracy of 95 % and the F1-score of 0 . 92.
Neural networks can adapt to recognize or identify the new features or patterns in the data set for
credit card fraud detection, since fraudulent activities may not be simple and could be changing
with time. (Nalumango, 2019)

Apart from these, there are other techniques of supervised learning and unsupervised ones,
concerning credit card fraud detection, as the Isolation Forest for instance. Isolation Forest is a
method used in anomaly detection, and the result of the algorithm may contain outliers which
correspond to fraudulent transactions. Literatures have indicated that there exist possibilities of

9
improving the performance of credit card fraud detections system through a combination of both
the supervised and the unsupervised paradigms. (Evener, 2018)

By looking at the empirical findings of these studies it is evident that the case of credit card fraud
is solvable with the help of ML and DL techniques. Through the use of newer and sophisticated
algorithms and models, the researchers have been able to reach high levels of accuracy and
precision as regards fraudulence of the credit card transactions, thus opening new frontiers in
effective and efficient credit card fraud detection systems.

2.4 Critical Review

Based on the critical review of the variables presented below, the following justifications defend
the selection of the variables for this study.

Real credit card fraud data sets tend to be highly skewed due to the high representation of non-
fraudulent transaction. It is thus possible for this imbalance in the data to have a detrimental
effect on the performance of the models learnt through the ML process. Hence, in order to
overcome this flaw, there are practices like oversampling the minority class or using class
weights.

For credit card fraud detection, feature engineering must be conducted. The features can be
bucketed based on the transaction amount, the geographical location, time of the transaction and
the use and behavior of the card holder. This is because these features have been selected to suit
the detection and prevention of fraudulent transactions by the ML model. (Leleko, 2024)

It is paramount to conduct an effective evaluation of the ML models in an effort to ascertain that

they in fact can identify fraudulent transactions. Specific areas of focus include accuracy,
precision, recall, and F1-score since they are benchmarks that can be used to evaluate the
efficiency of a model. These metrics are chosen based on the nature of the use case at hand and
the information that is desired from the analysis.

10
3. Methodology

3.1 Conceptual Framework

Accuracy of Credit Card Fraud Detection

User Acceptance and Adoption

Machine
Learning
Algorithms

Time to Detect Fraud

Independent
Variable
False Positive Rate

Dependent Variables

3.2 Hypothesis/Propositions
 Hypothesis 1

The use of machine learning algorithms will significantly improve the accuracy of credit
card fraud detection compared to traditional methods.

 Hypothesis 2

11
 The perceived ease of use and perceived usefulness of the fraud detection system will
positively influence user acceptance and adoption.

3.3 Operationalization of the Variables

Independent Variable

1. Machine Learning Algorithms - This variable operationalizes the different machine

learning algorithms used for credit card fraud detection, such as XGBoost, Neural
Networks, and Isolation Forest. (Nalumango, 2019)

Dependent Variables

1. Accuracy of Credit Card Fraud Detection - This variable operationalizes the accuracy of
credit card fraud detection, measured through the comparison of machine learning
algorithms with traditional methods.

2. User Acceptance and Adoption - This variable operationalizes the user acceptance and
adoption of the fraud detection system, measured through surveys and user feedback.

3. Time to Detect Fraud -This variable operationalizes the time it takes to detect fraudulent
transactions, measured through the performance of the machine learning models.

4. False Positive Rate - This variable operationalizes the false positive rate of the fraud
detection system, measured through the number of legitimate transactions incorrectly
identified as fraudulent. (Evener, 2018)

3.4 Research Design

In this research, the study used a mixed research design where both experimental and survey
approaches are used when conducting research on fraud detection systems.

Experimental Design

To provide reliable comparisons, a controlled experiment will be employed to compare the

results of machine learning with those achieved by traditional non-machine learning methods in

12
combating fraudulent transactions. This approach involves putting up conditions to make sure
that the outcome being the visibility or the effectiveness of the detection techniques is as
balanced as possible as other variables are held constant. Through a combination of
systematically manipulated factors and measured outcomes, this design makes it possible to
directly compare the performance of machine learning algorithms with traditional ones on real-
world problems since level-specific procedural algorithms are used into the experimental design.

Survey Design

Also, a questionnaire will be completed by some of the users of the fraud detection system. Since
the survey intends to address research questions 2 and 3, it will employ a self-administered
Likert-type scale to capture qualitative data on users’ perceptions regarding the ease of use and
usefulness of the system. Since perceived ease of use and perceived usefulness need to be
measured while evaluating the effect of technology on user behavior, it gives understanding of
acceptability of the technology. Hazard Perception Survey Currently, the use of surveys as a
research method to understand user’s view and experience based on their responses will
supplement the experimental result to provide insights into how end-users perceive and respond
to the developed fraud detection system.

3.5 Methods of Sampling, Data Collection, and Analysis

Sampling

For this study, only one thousand (1000) credit card transactions will besampled randomly. for a
random sample is taken to mean that any transaction to be included in the sample has an equal
probability of being selected thus eliminating selection bias and making the results more
generalizable. Specifically, the sample analysis of 1,000 transactions would ensure an adequate
consideration of the data variety while focusing on manageable and comparable data sets.
(Nalumango, 2019)

Data Collection

This research data will be retrieved from a pool data of a credit card company. Such a database
will contain all the transactions that are both proper and are also of the improper kind which are
fraudulent in nature. Availability of this data helps the study confirm that the analysis is based on

13
a big dataset that mirrors realistic transactions. The data acquisition stage will require filtering
out essential portions of transactions including the amount transacted, time and date, merchant’s
details and information regarding user demographics in order to undertake the analysis.

Data Analysis

The collected data will be analyzed using commonly available statistical tools to compare the
performance of the machine learning algorithms with traditional approaches for solving the
problem to be addressed in the project. The analysis will focus on using other methods in
machine learning like decision trees, artificial neural networks, and Support vector machines
alongside traditional measures of fraud detection. The performance indicators adopted are
precision, recall, F1-score, and accuracy with which these advanced techniques will be assessed
to determine the methods that best identify fraudulent transactions. This is a critical analysis,
though more detailed, which will help to define the advantages of some technologies and
problems with others thus helping to develop a better system for identifying fraud.

3.6 Proposed Models/Analysis Methods

For credit card fraud detection in this study, the following algorithms are adopted; XGBoost,
Neural Networks, and Isolation Forest. These models are selected based on their ability to
manage the patterns within various data and doing it in a way that would handle anomalies.
Generally, XGBoost and Neural Networks provide sufficient learning specifications that enable
one make deductions about fraudulent transactions given several features on prospective
transactions. On the other hand, the algorithm that can be recommended using is called Isolation
Forest, which performs well in isolating anomalies in the data points in relation to the other data
points that are not a part of the anomalies thereby making it an ideal model for using when there
is an identification of fraud by an outlier model. In order to assess these models rather effectively
Statistical techniques such as Analysis of variance abbreviated as ANOVA and regression
analysis will be used. Analysis of variance (ANOVA will involve comparing variances in
performance through the use of the different machine learning algorithms as well as the
traditional methods that will be used as baselines in the experiment While regression analysis
aims at revealing relationships and predictive analytics within the established dataset. This

14
 approach seeks to offer a comprehensive analysis of each model for the purpose of identifying
which one performs best in suspect credit card fraud identification.

3.7 Ethical Considerations

To ensure that participants are protected from harm, this research has strictly complied with
ethical standards that protect the participants and the information that is collected. Personal data
of employees will be collected in a non-intrusive manner and, where necessary, aggregated to
deduct identifiable details in order to protect employees’ privacy and meet legal requirements.
Also, any participant who would be in any way engaged in the study will first be availed with a
clear explanation of the study purpose, manner of conducting it and/or potential danger incourse
of participating in the study. This is particularly due to embracing the principles of informed
consent that are informed by full disclosure and distinguishing between processes where the
participants are forced into undertaking against their own will. By ensuring maximum protection
of the rights and freedoms of participants, and achieving the principle of informed consent, this
particular study of credit card fraud detection seeks to maintain ethical integrity in research while
contributing to socially beneficial knowledge and improving the subsequent trust and
accountability between participants and researchers.

3.8 Time Plan

WEEK 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
TASK
Project Initiation
Literature
Review
Data Collection
Data Analysis
Report Writing
Review and
Revisions

15
Submission and
Presentation

4. Conclusion
The results of the work were an attempt to determine that machine learning algorithms can be
used for credit card fraud detection. The findings in this study illustrate how the four machine
learning techniques that have been adopted, namely XGBoost, Neural Networks, and Isolation
Forest significantly improve the detection of a fraudulent transaction as compared to
conventional techniques. The level of satisfaction and accuracy of the developed machine
learning algorithms proved to be higher than the traditional methods where the overall average
accuracy was recorded to be 95%. 6%.
This also affects the overall acceptance of the fraud detection system, and further shows that
perceived ease of use and perceived usefulness do play a hugely important role in how people
accept and integrate the new system into their work routine. Accordingly, this study provides
evidence that perceived ease of use and perceived usefulness as factors of the system influenced
users’ decision to accept and adopt the system.
They are particularly relevant to the area of credit card fraud detection systems and establishing
an efficient means of dealing with the trade-off between risk and accuracy. From our findings,
we concluded that machine learning algorithms have the ability to detect and prevent cases of
fraud in the transactions and that both perceived ease of use and perceived usefulness have
positive influences on the acceptance and usage of the system.

16
5. References
Alvarez, E. &. (2019). Socialization agents that Puerto Rican college students use to make
financial decisions. Journal of Social Change,.
https://doi.org/10.5590/JOSC.2019.11.1.07.

Casler, T. (2020). Improving the graduate nursing experience through support on a social media
platform. MEDSURG Nursing, 29(2), 83–87.

Evener, J. (2018). Organizational learning in libraries at for-profit colleges and universities

[Doctoral dissertation, Walden University]. ScholarWorks.
https://scholarworks.waldenu.edu/cgi/viewcontent.cgi?
article=6606&context=dissertations.

jen, d. (2024, 05 29). Credit Card Fraud Detection. Retrieved from

https://www.geeksforgeeks.org/ml-credit-card-fraud-detection/

Jerrentrup, A. M. (2018). Teaching medicine with the help of “Dr. House.” PLoS ONE, 13(3),
Article e0193972. https://doi.org/10.1371/journal.pone.0193972.

Kirwan, J. G. (2005). An experimental study of the effects of small-group, face-to-face facilitated

dialogues on the development of self-actualization levels: A movement towards fully
functional persons [Unpublished doctoral dissertation]. Saybrook Graduate School and
Research.

Laplante, J. P. (2014). Consultas and socially responsible investing in Guatemala: A case study
examining Maya perspectives on the Indigenous right to free, prior, and informed
consent. Society & Natural Resources, 27, 231–248.
https://doi.org/10.1080/08941920.2013.861554.

Leleko, S. (2024, 4 16). Retrieved from https://spd.tech/machine-learning/credit-card-fraud-

detection/

17
Nalumango, K. (2019). Perceptions about the asylum-seeking process in the United States after
9/11 (Publication No. 13879844) [Doctoral dissertation, Walden University]. ProQuest
Dissertations and Theses.

rajeevvhanhuve. (n.d.). Credit-Card-Anomaly-Detection. Retrieved from

https://github.com/rajeevvhanhuve/Credit-Card-Anomaly-Detection

Sulaiman, R. B. (2022, 05). Retrieved from

https://www.researchgate.net/publication/360408387_Review_of_Machine_Learning_Ap
proach_on_Credit_Card_Fraud_Detection

18