ISSN: 2996-4954 Eswar Prasad G, et al.

(2023) 07
DOI: 10.47991/2996-4954/JCETAI-101

Research Article Journal of Contemporary Education Theory & Artificial Intelligence

Enhancing Performance of Financial Fraud Detection Through

Machine Learning Model
Authors:
Eswar Prasad Galla1*, Hemanth Kumar Gollangi2, Venkata Nagesh Boddapati3, Manikanth Sarisa4, Kiran
Polimetla5, Shravan Kumar Rajaram6, Mohit Surender Reddy7
1
Department of Computer Science, University of Central Missouri. Email: [email protected]
2
Department of Computer Science, Missouri State University. Email: [email protected]
3
Microsoft, Support Escalation Engineer. Email: [email protected]
4
Principal Software Engineer, Ally Financial Inc. Email: [email protected]
5
Adobe Inc, Software Engineer. Email: [email protected]
6
Microsoft, Support Escalation Engineer. Email: [email protected]
7
Microsoft, Support Escalation Engineer. Email: [email protected]
*
Corresponding author: Eswar Prasad Galla, Department of Computer Science, University of Central Missouri
Citation: Eswar Prasad G, Hemanth Kumar G, Venkata Nagesh B, Manikanth S, Kiran P, et al. (2023) Enhancing Performance of
Financial Fraud Detection Through Machine Learning Model. J Contemp Edu Theo Artific Intel: JCETAI-101.
Received Date: 10 October, 2023; Accepted Date: 18 October, 2023; Published Date: 23 October, 2023

Abstract
Despite attempts to reduce it, financial fraud continues to be a major problem in many industries, including healthcare, banking,
and insurance. Traditional fraud detection techniques, which are often manual, are inefficient, time-consuming, and costly. As
a result, methods that use AI and ML have been implemented to improve fraud detection procedures. This study examines the
application of ML algorithms for credit card fraud detection using a dataset consisting of 284,807 transactions made by
European cardholders in 2013, out of which 492 were fraudulent. Preprocessing steps, including Label Encoding, SMOTE for
handling class imbalance, and PCA for feature reduction, were applied to the dataset. On the training dataset have applied ML
based classification models like DT, SVM, and ANNs were employed to evaluate their performance. The models were assessed
using accuracy, precision, and recall as key metrics. The ANN model emerged as the best-performing model, achieving
98.41%precision, 98.69%accuracy, and 98.98%recall, outperforming both Decision Trees and SVM. This study highlights the
effectiveness of ML models, particularly ANNs, in improving financial fraud detection.
Keywords: Financial Fraud, Machine Learning, Credit Card Transaction Dataset, Detection.

Introduction networks, fraud detection in healthcare, banking, insurance, and

Financial fraud refers to the practice of acquiring money using mobile cellular networks, detection of anomalies in medical and
deceitful and illegal methods. Many different types of malware, and detection of anomalies in video surveillance. ML
businesses, banks, insurance companies, and government has many uses in the IoT realm, including location monitoring,
agencies are vulnerable to financial fraud [1]. Crimes against the detecting Android malware, automating the house, and
financial system, such as money laundering and fraudulent forecasting the incidence of heart disease[5].
financial transactions, have become more problematic for many
industries and companies in recent years. Financial fraud This project aims to investigate and evaluate ML methods for
continues to have a detrimental influence on society and the identifying instances of financial transaction fraud, with a
economy despite many efforts to reduce it. Every day, large specific emphasis on credit card fraud. With the rise of financial
amounts of money are wasted due to this crime [2]. A variety of fraud, traditional methods of detection, which are often manual,
methods for detecting fraud were developed long ago. Not only have proven to be time-consuming, costly, and inefficient. By
is physical labour required for most traditional operations, but it leveraging machine learning models, the study aims to identify
is also inefficient, costly, and prone to error. Despite their approaches that can effectively and accurately detect fraudulent
ineffectiveness, more research is being conducted to mitigate transactions. The objective of this study is to enhance financial
costs caused by fraudulent activities. The advancement of AI has sector fraud detection systems and reduce economic losses by
allowed for the use of data mining and ML to detect fraudulent comparing several models, such as DT, NB, and ANN, and
activities in the financial sector. Both supervised and highlighting the advantages and disadvantages of each
uncontrolled methods were used to anticipate fraud acts. approach. The following are the main contribution of the study
Classification is the gold standard method for detecting financial on financial fraud detection based on Credit card transaction
fraud[3]. dataset:
• By comparing models like DT, NB, and ANNs, the study
ML is a branch of AI that enables machines to draw conclusions
identifies the most effective techniques for fraud detection
from past data by examining patterns. With Google DeepMind's
based on key metrics like recall, precision, and accuracy.
phenomenal achievement in 2015, AI and ML have reached new
heights[4]. A few real-world applications of deep anomaly • The study applies its analysis on a real-world credit card
detection include security breach detection in computer transaction dataset, which adds practical relevance to its

J Contemp Edu Theo Artific Intel, 2023 ISSN: 2996-4954 Page: 1 of 7

 Citation: Eswar Prasad G, Hemanth Kumar G, Venkata Nagesh B, Manikanth S, Kiran P, et al. (2023) Enhancing Performance of
Financial Fraud Detection Through Machine Learning Model. J Contemp Edu Theo Artific Intel: JCETAI-101.

findings on model performance and effectiveness in 85% with an accuracy of above 90%. According to our research,
detecting fraudulent activities. the tiered random forest achieves a recall of 72% and a precision
• It demonstrates how preprocessing techniques like SMOTE of 85%, making it the most effective algorithm compared to
for handling class imbalance and PCA for feature reduction SVM and logistic regression [9].
can improve model performance in fraud detection.
• The study shows that Artificial Neural Networks (ANNs) In this paper Erfani, Shoeleh and Ghorbani, (2020), provide a
outperform other models according to accuracy, precision, streamlined system for identifying fraudulent activities. In order
and recall, making it a highly effective technique for to identify fraud, our methodology employs deep support vector
financial fraud detection. data description after a unique preprocessing and subsampling
phase. They offer a trend analysis that takes into account the
Structure of the paper dimensions of the training and test datasets as well as the
This research is organised in the following way: Predicting model's performance as measured by ROC-AUC and AP. Last
online sales is the focus of Section 2, which summarises current but not least, our method beats the advance binary classifiers,
approaches. The approach, including data management and RF and SVM, in several tests. The best values are 90% for AP
model application, is described in Section 3. The outcomes of and 93% for ROC-AUC, demonstrating its outstanding
the experiments are detailed and discussed in Section 4. Section performance [10].
5 presents the important findings and suggests areas for further
research. This paper Arun and Venkatachalapathy, (2020), announces a
new C-LSTM model for detecting credit fraud that is based on
Literature Review DL. Two steps are involved in the suggested C-LSTM model:
This section reviews key machine learning research on similar preprocessing and classification. Using a German Credit and
datasets and challenges, highlighting influential methods and Kaggle's CCFD datasets, we verify a performance of a C-LSTM
studies. Table 1 summarizes the relevant literature for financial model. The obtained experimental outcomes demonstrated that,
fraud detection. when applied to a German credit and CCFD dataset, the C-
LSTM model performed well with accuracy of 94% and 94.65%
This paper Rai and Dwivedi, (2020), proposes a way to detect
[11].
fraudulent activity in credit card data by using a NN based
unsupervised learning methodology. This new approach
Table 1: Comparative research of Financial Fraud
outperforms the state-of-the-art AE, LOF, IF, and K-Means
Detection using machine and deep learning techniques
clustering algorithms. In comparison to the existing approaches,
Ref Methodolog Datase Result Limitation and
which include AE, IF, LOF, and K Means, the proposed NN-
y t future work
based fraud detection system achieves an accuracy of 99.87%
[6]. [1] Neural Credit 99.87% May not
Network card accurac generalize to
(NN) based data y different types
This study Hidayattullah, Surjandari and Laoh, (2020), uses a
unsupervised of fraud or data
variety of ML methods grounded on meta-heuristic optimisation
learning variations
to construct reliable financial statement fraud prediction models.
technique
Two different types of classification algorithms were employed:
SVMs and Back Propagation Neural Networks. This study's top [2] Meta- Financ SVM May be limited
classifier is a SVM, with 96.15% accuracy achieved by heuristic ial with by the quality
optimising its parameters using a Genetic Algorithm [7]. optimization statem Genetic and
with Back ents Algorit representativene
This paper Mubalaike and Adali, (2018), seeks to comprehend Propagation hm: ss of financial
the ways in which DL models may be helpful in accurately Neural 96.15% data used
identifying fraudulent transactions. The preprocessed data is Networks accurac
then subjected to the best ML and DL methods, including and SVM y
ensembles of decision trees (EDTs) and SAEs and RBM [3] Ensemble of One EDT: Performance
classifiers. An optimum accuracy value are 90.49%, 80.52%, Decision month 90.49%, might vary with
and 90.49%, respectively. A closer look at the findings shows Tree (EDT), of SAE: different
that RBM outperforms the alternatives [8]. Stacked financi 80.52%, datasets and the
Auto- al logs RBM: extensive
This study Gardner et al. (2019), stress the need of creating a Encoders from a 91.53% computation
system that can identify anomalies in financial transactions (SAE), RBM mobile required
using three different components. In order to build the system, money
many RFC with distinct fitness functions are fine-tuned. By service
optimising the RF parameters to meet the fitness function, the [4] Three-tiered Not 96% High
procedure is carried out using a randomised grid search. When anomaly specifi correct complexity of
all of the models are finished, they are compared to create three detection ed fraud tuning and
levels of discovered frauds, with varying degrees of accuracy in system with classific parameter
each level. Detected frauds may be categorised into multiple randomized ation; optimization;
levels for improved recall and precision. Using this method, we grid search Precisio may not scale
are able to accurately classify 96% of frauds while detecting and multiple n over well

J Contemp Edu Theo Artific Intel, 2023 ISSN: 2996-4954 Page: 2 of 7

 Random 90% for
Forest 85% of
classifiers detected
frauds
[5] Preprocessin Not AP: Framework's
g and specifi 90%, performance
subsampling ed ROC- might depend
with Deep AUC: on the specifics
Support 93% of preprocessing
Vector Data and
Description subsampling
steps
[6] C-LSTM Germa German May require
model n Credit: substantial
Credit 94%, computational
dataset Kaggle resources and
, Credit may not handle
Kaggle Card all fraud types
Credit Fraud equally
Card Detecti
Fraud on:
Detecti 94.65%
on
dataset

Research gaps
While existing studies on fraud detection have demonstrated
significant advancements with various methodologies—such as Figure 1: Proposed flowchart of financial fraud detection
NN-based unsupervised learning, meta-heuristic optimization, system.
and deep learning techniques—there remains a notable research A following data flowchart of financial fraud detection system
gap in generalizing these methods across diverse datasets and steps are listed, shows in figure 1. Each level of data processing
real-world scenarios. Many approaches are optimized for in the system is explained in depth.
specific datasets, such as mobile money transactions or financial 1) Data collection
statements, limiting their applicability to broader contexts. The process of data collection involves collecting relevant
Additionally, the complexity of tuning and preprocessing information from many sources, such as sensors and databases,
methods, along with the variability in performance metrics like in order to construct a large and representative dataset for
accuracy and precision, indicates a need for more robust and analysis. There is a notable disparity between the 284,807 credit
adaptable frameworks. Future research should focus on card transactions recorded for European cardholders in
developing universal models that integrate advanced techniques September 2013 and the 492 fraudulent transactions.
and improve generalization, while also addressing the resource- 2) Data preprocessing
intensive nature of current optimization processes. The term "data processing" refers to the steps used to make raw
data more suitable for analysis. This includes resolving missing
Research Methodology data, removing duplicates, and encoding category variables.
In this research aim to provide the efficient ML based financial Data preparation for modelling involves checking for quality
fraud detection system. Beginning with the gathering of a and consistency in order to boost the efficiency of later analyses
dataset consisting of 284,807 transactions by European cards, and ML algorithms. Here are some of the most important pre-
including 492 instances of fraud, the technique for evaluating processing steps:
this information for the purpose of detecting fraud uses a
• Label Encoding: Label Encoding is a way to make
structured approach. The data is carefully prepared for analysis
numerical values out of category data. Algorithms that work
by filling in missing values, eliminating duplicates,
with numerical input can handle categorical data since each
standardising it, and encoding categorical variables employing
category is given its own distinct integer.
Label Encoding. Using SMOTE and PCA for feature selection
• SMOTE: The SMOTE, which generates synthetic samples
helps with class imbalance. The dataset is then divided into
for the minority class, is one way to remedy class
training (80%) and testing (20%) subsets. An assortment of
imbalance. By interpolating between preexisting data
classification models, such as DT, NB, and ANNs, are tested and
points, it generates additional, synthetic data points, hence
assessed for their accuracy and efficacy in differentiating
enhancing the performance of ML models on unbalanced
between genuine and fraudulent transactions. Figure 1 shows a
datasets.
flow diagram of the system that detects financial fraud.
• Feature Selection: Principle Component Analysis converts
a dataset's characteristics into a new collection of variables
known as principle components, therefore lowering a
number of features in the dataset. These elements capture
the majority of the volatility in the data, enabling a more
condensed depiction without sacrificing crucial details.

J Contemp Edu Theo Artific Intel, 2023 ISSN: 2996-4954 Page: 3 of 7

 Citation: Eswar Prasad G, Hemanth Kumar G, Venkata Nagesh B, Manikanth S, Kiran P, et al. (2023) Enhancing Performance of
Financial Fraud Detection Through Machine Learning Model. J Contemp Edu Theo Artific Intel: JCETAI-101.

3) Data Splitting accuracy, and precision. A dataset, performance metrics, ANN

A potential solution to address class imbalance is the SMOTE, model output, and comparative analysis are detailed in the parts
which creates synthetic samples for the minority class. The that follow.
effectiveness of ML models on imbalanced datasets is improved
by interpolating between extant data points to create new, Dataset Description
synthetic data points. The model will be trained using 80% of The dataset is extremely skewed since it includes fraudulent
the data and tested with 20%. This ratio is intended to ensure transactions (492 out of 284,807) from European cardholders in
that a smaller sample size is used for testing once the model has September 2013. PCA was used to convert the majority of the
been trained on the majority of the data. features in order to preserve privacy and secrecy. This produced
4) Classification Models 28 PCA-transformed numeric values (V1 through V28). The
This section describes how the Credit Card Transaction dataset "Time" and "Amount" characteristics are the only ones that have
was analysed using several categorisation techniques. These not changed. This dataset is great for studies on financial fraud
models are used to assess and compare their performance across detection since the goal variable "Class" displays whether a
the provided features. The goal of this analysis is to find the transaction is fraudulent (1) or regular (0). The figure 2 shows
model that provides the most insightful and accurate findings the bar plot of (a) shows the total number of counts of each class
when applied to product ratings and customer reviews. and (b) shows the after SMOTE balanced approx. count plot.
a) Decision Tree
Regression and classification problems are two of the most
common and extensively utilised supervised ML techniques that
are available: decision trees. The Decision Tree method has a
simple yet very effective intuition. Classifying non-linearly
separable data is possible with little effort spent training the
algorithm. When contrasted with KNN and other classification
algorithms, it demonstrates remarkable speed and efficiency.
The two most used metrics for selecting attributes are entropy
and information gain. Figure 2: Count plot of after data balancing and before.
b) SVM
SVMs are another well-liked supervised learning for
classification method. This classifier uses a hyper plane to
divide the dataset into categories; it's discriminative. One way
to divide up n-dimensional data is via a hyper plane. A hyper
plane that maximises the margin is chosen from among
numerous possible separations of the dataset. When compared
to other algorithms, SVM uses less processing time while yet
providing good accuracy.
c) ANN
The three layers that make up an ANN are an input, hidden, and
output layers. An input layer is responsible for supplying the
neural network with input information; a hidden layer is a brain
of a network, constantly updating the weights to improve
performance; and an output layer is where a network's results
Figure 3: Correlation matrix of credit card dataset.
are rendered in class terms. The propagation function and
learning rule determine the neural network's output. Equation The associations between the many characteristics seen in
(1) expresses the propagation function, which controls the inputs Figure 3, such as "Time," "Amount," and anonymized variables
to the jet neuron from the outputs of the previous neurons. "V1" to "V28," are represented visually by the credit card
dataset's correlation matrix. The correlation coefficient among
𝑃𝑡 (𝑡) = ∑𝑖 0𝑖 (𝑡) × 𝑤𝑖𝑗 + 𝑏……. (1) two variables is shown in each matrix cell, with colour intensity
denoting the association's strength and direction. Shades of blue
where the previous neuron's output is represented by 𝑂𝑖(𝑡), the indicate significant negative connections, white or lighter
propagation function is represented by 𝑃𝑗 (𝑡), the weight is colours show little to no link, and red hues indicate strong
represented by 𝑤𝑖𝑗, and the bias is represented by 𝑏. positive correlations. The diagonal line from the top left to the
It is possible to train a NN to do well with a certain set of inputs bottom right is solid red, indicating perfect positive correlation
by adjusting its parameters according to a learning rule. As it as each variable is compared with itself. This matrix is crucial
learns, the network's weights are adjusted to improve output for identifying patterns, relationships, and potential anomalies
computation according to the learning rule[12]. in the dataset, aiding in effective data analysis and decision-
making.
Result Analysis and Discussion
Here we present the experimental outcomes of ML models Performance Measures
trained on the Kaggle dataset using the Python simulation tool Four separate criteria were utilized to evaluate each model's
with the purpose of detecting financial fraud. For the purpose of performance: F1-score, recall, accuracy, and precision. These
assessing ML models' efficacy using a confusion matrix, recall, evaluation parameters are detailed below:

J Contemp Edu Theo Artific Intel, 2023 ISSN: 2996-4954 Page: 4 of 7

 Citation: Eswar Prasad G, Hemanth Kumar G, Venkata Nagesh B, Manikanth S, Kiran P, et al. (2023) Enhancing Performance of
Financial Fraud Detection Through Machine Learning Model. J Contemp Edu Theo Artific Intel: JCETAI-101.
1. Confusion Matrix Performance of ANN model for financial fraud detection
Confusion matrices, sometimes known as error matrices, may
be used to display the performance of algorithms. The structure
is a table. The matrix's rows show actual instances of the classes, 99.1 98.98
while its columns show anticipated instances, or vice versa. It is 99
98.9
used to calculate F1 scores, recall, precision, and accuracy [13]. 98.8 98.69
The confusion matrix makes use of the following terms: 98.7

in%
98.6 98.41
• True Positive (TP): This situation occurs when the 98.5
98.4
expected and actual classes of a data point are both 1. 98.3
98.2
• True Negative (TN): A data point is considered to have 98.1
this property when its anticipated and actual classes are both Accuracy Precision Recall
0.
• False Positive (FP): This happens when a data point has a
Performance measures
predicted class of 1 but a real class of 0.
• False Negative (FN): To put it simply, this happens when
a data point has a real class of 1 but an anticipated class of
0. Figure 4: Financial Fraud Detection performance with ANN

2. Accuracy The figure 4 shows the financial fraud detection performance

A number of optimistic forecasts that are really right is measured with ANN model according to accuracy of 98.69%, precision of
by accuracy. As seen in equation (2), genuine positive / total 98.41% and recall of 98.98%, respectively.
projected positive is the ratio that determines precision in
imbalanced classification:
𝑇𝑃 + 𝑇𝑁
𝐴𝑐𝑢𝑟𝑎𝑐𝑦 = … . (2)
𝑇𝑃 + 𝑇𝑁 + 𝐹𝑁 + 𝐹𝑃
3. Precision
Precision measures the effectiveness of a classifier in predicting
positive samples. Simply divide the sum of all positive
predictions by the number of actual positive samples, as shown
in Equation (3), and you have the answer:
𝑇𝑃
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = … . (3)
𝑇𝑃 + 𝐹𝑃
4. Recall
The test result will accurately identify fraudulent transactions if
recall is high. Recall, often called sensitivity, is defined as a ratio
of TP to total real positives, as seen in equation (4): Figure 5: Confusion matrix for ANN Model.
𝑇𝑃
𝑅𝑒𝑐𝑎𝑙𝑙 = … . (4)
𝑇𝑃 + 𝐹𝑁 Figure 5 shows the ANN model's confusion matrix, which
According to these measures evaluate a performance of ML clearly shows how well the model performed in detecting fraud.
models for financial fraud detection. There were 55,752 cases of fraud (TP) and 55,983 cases of non-
fraud (TN) that the model accurately recognised. However, it
Experiment Results also made 597 FP errors, where non-fraudulent transactions
An experimental outcome of several ML and DL models applied were incorrectly flagged as fraud, and 2,394 false negative
to the Credit Card Transaction dataset are displays in this errors, where fraudulent transactions were missed. This matrix
section. An outcome is displayed using figures, graphs, and is crucial for evaluating the model’s accuracy and reliability,
tables, providing a detailed overview of each model's highlighting areas where it performs well and where
performance. improvements are needed.

Comparative analysis
The table 2 below provides a comparative analysis of multiple
DL and ML models used for financial fraud detection,
specifically applied to a Credit Card Transaction dataset. For a
thorough evaluation of the models' efficacy in identifying
fraudulent behaviour in this domain, it details their performance
indicators.

J Contemp Edu Theo Artific Intel, 2023 ISSN: 2996-4954 Page: 5 of 7

 Citation: Eswar Prasad G, Hemanth Kumar G, Venkata Nagesh B, Manikanth S, Kiran P, et al. (2023) Enhancing Performance of
Financial Fraud Detection Through Machine Learning Model. J Contemp Edu Theo Artific Intel: JCETAI-101.
Table 2: Comparison between various model for enhancing the
performance of Financial Fraud Detection. Comparison of different models of Recall

Models Accuracy Precision Recall 96.4 98.98

100 81
Decision Tree 88 95 81
[14]

IN %
SVM [15] 72.3 60 96.4 50
ANN 98.69 98.41 98.98
0
Decision SVM ANN
Comparison of different models of Tree
Accuracy MODELS

Figure 8: Comparison of different models of Recall.

IN %

88 98.69
72.3 The three models' recall scores are compared in Figure 8. The
model that performs best in terms of memory is the ANN, which
stands out with the greatest recall score of 98.98%. The
DECISION TREE SVM ANN DT model, on the other hand, has the lowest recall, at 81%.
MODELS
This study's results show that when compared to more
conventional ML models like DT and SVM, the ANN model
performs far better in detecting financial fraud. The confusion
Figure 6: Comparison of different models of Accuracy matrix shows that the ANN achieves an outstanding
accuracy98.69%, precision98.41%, and recall98.98% when it
Figure 6 is a bar graph that shows how various models' accuracy comes to properly recognising both fraudulent and non-
is compared. The graph indicates that the ANN achieves the fraudulent transactions. The comparative analysis highlights the
highest accuracy with a score of 98.69%, while the SVM model ANN's significant advantages over the SVM, which exhibited
exhibits the lowest accuracy, scoring 72.3%. the lowest accuracy (72.3%) and precision (60%), and the
Decision Tree model, which, despite better precision (95%),
C ompar iso n of dif f e r e nt mode ls of lagged in recall (81%). These findings indicate that the ANN not
Pr e c ision only minimizes false positives and negatives effectively but also
ensures a balanced performance across key metrics, suggesting
120 that deep learning approaches are more adept at handling
95 98.41
100 imbalanced datasets, making them particularly suitable for
financial fraud detection applications.
80
60 Conclusion and Future Work
IN %

60
Frauds are said to be dynamic and lacking in patterns, making
40 them difficult to identify. Fraudsters profit from new
technological developments. They manage to circumvent all
20 security measures, which leads to a massive financial loss. One
0
way to keep tabs on fraudulent transactions is to use data mining
Precision techniques to analyze and detect unusual behaviors. Credit card
transaction analysis has been the primary focus of this study's
Decision Tree SVM ANN investigation into ML methods for financial fraud detection.
ML's efficacy in differentiating between genuine and fraudulent
Figure 7: Comparison of different models of Precision. transactions was proved via the application of classification
models such as DT, SVM, and ANN. The most resilient model
Figure 7 illustrates the precision comparison among the models. for fraud detection was ANN, which had the greatest
The bar graph shows that the SVM model has the lowest performance among the models with accuracy (98.69), precision
precision score at 60%, while the ANN attains a highest (98.41), and recall (98.98). The findings point to the value of
precision with a value of 98.41%, respectively. ML for detecting financial crime and provide a way forward for
creating better fraud prevention systems.

Future research can focus on incorporating more sophisticated

DL models, like CNN or RNN, to further improve fraud
detection performance. Additionally, hybrid models that
combine various machine learning techniques may provide
better accuracy in identifying fraudulent transactions. Another
important aspect for future investigation is the application of
real-time detection systems for fraud prevention, as well as

J Contemp Edu Theo Artific Intel, 2023 ISSN: 2996-4954 Page: 6 of 7

 Citation: Eswar Prasad G, Hemanth Kumar G, Venkata Nagesh B, Manikanth S, Kiran P, et al. (2023) Enhancing Performance of
Financial Fraud Detection Through Machine Learning Model. J Contemp Edu Theo Artific Intel: JCETAI-101.

exploring methods for reducing false positives, which are and Information Security, IWBIS 2020, 2020. doi:
critical in practical implementations. Model generalizability 10.1109/IWBIS50925.2020.9255563.
may be further improved by using more and more varied 8. A. M. Mubalaike and E. Adali, “Deep Learning Approach
datasets. for Intelligent Financial Fraud Detection System,” in
UBMK 2018 - 3rd International Conference on Computer
References Science and Engineering, 2018. doi:
1. A. Mousa, “Detecting Financial Fraud Using Data Mining 10.1109/UBMK.2018.8566574.
Techniques: A Decade Review from 2004 to 2015,” J. Data 9. C. Gardner, D. C.-T. Lo, J.-C. Chern, P. Paschos, and C.
Sci., vol. 14, no. 3, pp. 553–570, 2016, doi: Ng, “Tiered Financial Fraud Detection Utilizing Precision
10.6339/jds.201607_14(3).0010. Stratified Random Forest Assembly,” in 2019 IEEE 5th
2. N. F. Ryman-Tubb, P. Krause, and W. Garn, “How International Conference on Big Data Intelligence and
Artificial Intelligence and machine learning research Computing (DATACOM), 2019, pp. 254–257. doi:
impacts payment card fraud detection: A survey and 10.1109/DataCom.2019.00047.
industry benchmark,” Engineering Applications of 10. M. Erfani, F. Shoeleh, and A. A. Ghorbani, “Financial
Artificial Intelligence. 2018. doi: Fraud Detection using Deep Support Vector Data
10.1016/j.engappai.2018.07.008. Description,” in Proceedings - 2020 IEEE International
3. M. R. Kishore Mullangi, Vamsi Krishna Yarlagadda, Conference on Big Data, Big Data 2020, 2020. doi:
Niravkumar Dhameliya, “Integrating AI and Reciprocal 10.1109/BigData50022.2020.9378256.
Symmetry in Financial Management: A Pathway to 11. G. K. Arun and K. Venkatachalapathy, “Convolutional
Enhanced Decision-Making,” Int. J. Reciprocal Symmetry Long Short Term Memory Model for Credit Card
Theor. Phys., vol. 5, no. 1, pp. 42–52, 2018. Detection,” in Proceedings of the 4th International
4. S. K. R. A. Sai Charan Reddy Vennapusa, Takudzwa Conference on Electronics, Communication and Aerospace
Fadziso, Dipakkumar Kanubhai Sachani, Vamsi Krishna Technology, ICECA 2020, 2020. doi:
Yarlagadda, “Cryptocurrency-Based Loyalty Programs for 10.1109/ICECA49313.2020.9297606.
Enhanced Customer Engagement,” Technol. Manag. Rev., 12. A. Khashman and K. Dimililer, “Neural networks
vol. 3, no. 1, pp. 46–62, 2018. arbitration for optimum DCT image compression,” in
5. H. Zhou et al., “A distributed approach of big data mining EUROCON 2007 - The International Conference on
for financial fraud detection in a supply chain,” Comput. Computer as a Tool, 2007. doi:
Mater. Contin., 2020, doi: 10.32604/CMC.2020.09834. 10.1109/EURCON.2007.4400236.
6. A. K. Rai and R. K. Dwivedi, “Fraud Detection in Credit 13. H. M and S. M.N, “A Review on Evaluation Metrics for
Card Data using Unsupervised Machine Learning Based Data Classification Evaluations,” Int. J. Data Min. Knowl.
Scheme,” in Proceedings of the International Conference Manag. Process, 2015, doi: 10.5121/ijdkp.2015.5201.
on Electronics and Sustainable Communication Systems, 14. K. Kumain, “Analysis of Fraud Detection on Credit Cards
ICESC 2020, 2020. doi: using Data Mining Techniques,” Turkish J. Comput. Math.
10.1109/ICESC48915.2020.9155615. Educ., 2020, doi: 10.17762/turcomat.v11i1.13590.
7. S. Hidayattullah, I. Surjandari, and E. Laoh, “Financial 15. D. Zhang, B. Bhandari, and D. Black, “Credit Card Fraud
statement fraud detection in indonesia listed companies Detection Using Weighted Support Vector Machine,” Appl.
using machine learning based on meta-heuristic Math., vol. 11, no. 12, pp. 1275–1291, 2020, doi:
optimization,” in 2020 International Workshop on Big Data 10.4236/am.2020.1112087.

Copyright: © 2023 Eswar Prasad G. This Open

Access Article is licensed under a Creative
Commons Attribution 4.0 International (CC BY
4.0), which permits unrestricted use, distribution,
and reproduction in any medium, provided the
original author and source are credited.

J Contemp Edu Theo Artific Intel, 2023 ISSN: 2996-4954 Page: 7 of 7