See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/378104413

Fraud_Detection_ML: Machine Learning Based on Online Payment Fraud

Detection

Article in Journal of Computing and Communication · February 2024

DOI: 10.21608/jocc.2024.339929

CITATIONS READS

0 2,582

11 authors, including:

Nashwa Shaker Ragab Dr-Diaa Salama

Alamein International University Misr International University
10 PUBLICATIONS 32 CITATIONS 185 PUBLICATIONS 2,627 CITATIONS

SEE PROFILE SEE PROFILE

Omnia Elrashidy Omar Adel

Alamein International University Ahram Canadian University
7 PUBLICATIONS 2 CITATIONS 3 PUBLICATIONS 16 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Omnia Elrashidy on 09 February 2024.

The user has requested enhancement of the downloaded file.

 Journal of Computing and Communication Vol.3 , No.1 , PP. 116-131 , 2024

Fraud_Detection_ML: Machine Learning Based on Online

Payment Fraud Detection
Maged Farouk a, Nashwa S Ragaba, Diaa Salama*b,c, Omnia Elrashidy a, Nada Ghorab a, Jevana Hany a, Alaa
Amr a, Omar Adel a, Kriols Saad a, Khaled Ali a, Reda Elazaba
a Department of Business Information Systems, Faculty of Business, Alamein University, Alamein, Egypt
b Faculty of Computers Science, Misr International University, Cairo, Egypt
c, Faculty of Computers and Artificial Intelligence, Benha University, Benha, Egypt
*
Corresponding Author: Diaa Salama [[email protected]]

ARTICLE DATA ABSTRACT

Online payment fraud detection is crucial for safeguarding e-commerce transactions against
Article history: sophisticated fraudsters who exploit system vulnerabilities. This paper proposes an efficient
Received 08 Jan 2024
framework for predicting online payment fraud, employing six diverse machine learning
Revised 14 Jab 2024
Accepted 04 Feb 2024 algorithms, namely constant, CN7Rule induction, KNN, Tree, Random Forest, Gradient boosting,
Available online SVM, Logistic regression, Naive Bayes, Ada boost, Neural network, and stochastic gradient
descent, on three distinct datasets. The gradient-boosting algorithm consistently outperformed
Keywords: others through rigorous testing, achieving an impressive accuracy rate of 99.7%. This algorithm
Online payment fraud, demonstrated resilience across various testing scenarios, establishing itself as the most effective
Machine-Learning, online payment fraud detection solution. With the highest accuracy score of 99.7% in all testing
gradient boosting,
phases, gradient boosting is optimal for preemptive measures against fraudulent activities in
CN2Rule Induction,
electronic transactions, providing a robust defense mechanism for e-commerce platforms.
fraud deduction

1. Introduction
Online payment fraud detection is a process that prevents fraudulent activities in online transactions. It
involves device fingerprinting, geolocation, behavioral analysis, transaction tracking, and two-factor
authentication. Machine learning and AI algorithms continuously adapt to new fraud strategies, and
cooperation between payment service providers and financial institutions is beneficial [1].
Online payment fraud is a problem that arises from dishonest and illegal actions taken during electronic
transactions. Unauthorized transactions, identity theft, compromised payment credentials, phishing, social
engineering, insufficient security protocols, account takeover, difficulties with cross-border transactions,
and risks associated with developing technologies are some of the major problems [2].
Machine learning is an evolving branch of computational algorithms designed to emulate human
intelligence by learning from the surrounding environment. They are considered the working horse in the
new era of big data. Techniques based on machine learning have been applied successfully in diverse fields
ranging from pattern recognition, computer vision, spacecraft engineering, finance, entertainment, and
computational biology to biomedical and medical applications [3].
Machine learning plays a crucial role in addressing the challenge of online payment fraud by enabling
automated, data-driven fraud detection and prevention systems. Here's how machine learning is applied to
combat online payment fraud [4].
The main contribution of this paper follows: we use six algorithms, and we made predictions for online
payment fraud; we use cross-validation (10), training (80%), and testing (20), and the best algorithm was
gradient boosting with Accuracy (0.997).
The rest of the paper can be organized as follows: Machine learning is an effective technique when it
comes to identifying and stopping online payment fraud. It can analyze vast data volumes, spot trends, and
116
 Diaa Salama et al. Journal of Computing and Communication Vol.3 , No.1 , PP. 116-131 , 2024

generate precise forecasts. Machine learning models can recognize suspicious trends and flag them by
utilizing features like transaction amounts, locations, timestamps, user behavior, and device information.
Human and tantalite ending Caprice is vital. The validity and applicability of training data were the focus
of the Machine Emin Motel lawsuits.

Fig1.Types of Common Fraud

2. Related Work

In [5], the authors explained that the model resulted in a considerable reduction in fraud and savings of
101,970.52 EGP out of 131,297.83 EGP. It was constructed using the IBM SPSS modeler's decision tree.
It obtained an impressive 88.45% accuracy and 93.5% precision. Plotting to increase from an anticipated
$10.7 billion in 2015 to $25.6 billion by the end of the decade, online and mobile fraud will significantly
influence the worldwide e-payments business.
In [6], the authors explained that the model is tested against the Random Forest and Gradient Boosting
Machine algorithms to determine its efficacy. Findings demonstrate the Light Gradient Boosting Machine's
strong performance; in real datasets, it achieved a total recall rate of 99% and offered prompt feedback.
This demonstrates how well the model detects credit card fraud.
In [7], the authors explained the process of detecting payment fraud. For this purpose, machine learning
classifiers such as Bagging Ensemble Learner, C4.5 decision trees, and Naïve Bayes are suggested. These
classifiers' performance is measured using evaluation measures such as Accuracy, recall rate, and precision-
recall curve area rate. Three thousand two hundred ninety-three fraudulent transactions were included in
the dataset, which included approximately 297,000 credit card transactions from September 2013 to
November 2017. Outstanding performance is shown by machine learning classifiers, which have a
precision-recall curve ratio between 99.9% and 100%. With an astounding 94.12% accuracy rate in
predicting fraudulent transactions, C4.5 decision trees are the most successful classifier.
In [8] the authors explained for these detection methods, assessment criteria include specificity,
Accuracy, sensitivity, and precision. Accuracy rates for Naive Bayes, K-Nearest Neighbor, Support Vector
Machine, and logistic regression are 97.53%, 97.53%, 94.98%, and 99.51%, respectively. The comparative
results of the study show that logistic regression is the best algorithm out of these. Unlike Naive Bayes, K-
Nearest Neighbor, and Support Vector Machine, logistic regression exhibits optimal Accuracy. These
results highlight the superiority of logistic regression over alternative methods in identifying credit card
fraud.
In [9], the authors explained the study investigates Fraud Detection Systems (FDS) for credit cards using
naïve Bayes, support vector machines, random forests, decision trees, OneR, and AdaBoost machine
learning approaches. A dataset is evaluated using a variety of machine learning approaches, with an
117
 Diaa Salama et al. Journal of Computing and Communication Vol.3 , No.1 , PP. 116-131 , 2024

emphasis on Accuracy, to produce performance measures. The study concludes that the random forest
classifier performs better than all the other techniques examined.
In [10] the authors explained the primary goal of this research is to study machine learning methods.
The Ada boost algorithm and the Random Forest algorithm are the algorithms that are employed. Outcomes
from both Accuracy, precision, recall, and F1-score serve as the foundation for algorithms. The confusion
matrix serves as the basis for plotting the ROC curve. When comparing the Random Forest and Ada boost
algorithms, the method with the highest Accuracy, precision, recall, and F1 score is deemed the most
effective for fraud detection.

3. Methodology
3.1Datasets Descriptions
The first dataset consists of 10 features, and it has 1,048,576 records. The dataset was split into two
partitions: 80% for training and 20% for testing. Below is a comprehensive description of each feature.
Step: An interval of time equal to one hour. Type: Indicates the kind or classification of the virtual
transaction. Amount: Indicates how much money was exchanged in this transaction. NameOrig: Indicates
which client started the transaction. OldbalanceOrg: This shows the customer's balance before the
transaction. NewbalanceOrig: Shows the customer's balance following the transaction. NameDest: Indicates
who will receive the transaction. OldbalanceDest: Stores the recipient's starting balance before the
transaction. NewbalanceDest: These variables record the recipient's new balance after the transaction.
IsFraud: This indicates if the transaction is thought to be fraudulent or not.

TABLE I
FEATURES OF DATASET

Features Type Value

step Numerical From 1 to 743
Type Classification Payment or transfer or debit. etc
amount Numerical 0 to 92.4 m
nameOrig Alphanumerical String
oldbalanceOrg Numerical 0 to 59.6m
newbalanceOrig Numerical 0 to 49.6m
nameDest Alphanumerical String
oldbalanceDest Numerical 0 to 356 m
newbalanceDest Numerical 0 to 356 m
isFraud Classification 0 or 1

Alphanumeric values consist of a combination of letters and numbers. This current investigation, which
involved 12 undergraduates, demonstrated that angular orientation had little effect on the delay in
determining whether a disoriented character was a letter or a digit [11].

3.2 Used Algorithms

These datasets were fed into twelve distinct machine learning algorithms: Gradient Boosting, K Nearest
Neighbour (k-NN), and Logistic Regression. Random Forest, Decision Tree, Constant, CN7 Rule induction,
SVM, Ada boost, neural network, stochastic gradient descent, and Naive Bayes algorithm. Statistics,
including Accuracy, recall, precision, and MCC, were produced for each of the algorithms. Next, a chart
and a comparison were made of the results. Further in the paper are the results, graphics, and a discussion.

118
 Diaa Salama et al. Journal of Computing and Communication Vol.3 , No.1 , PP. 116-131 , 2024

1-Gradient Boosting: A powerful family of machine-learning algorithms known as gradient boosting

machines has demonstrated notable effectiveness in various real-world applications. They can be learned
in relation to various loss functions, for example, and are highly customizable to the application's specific
requirements [12].
2-(k-NN): k-nearest neighbor (KNN) is one of the most prominent, simple, and basic algorithms used in
machine learning and data mining. However, KNN has limited prediction ability, i.e., KNN cannot predict
any instance correctly if it does not belong to any predefined classes in the training data set [13].

3-logistic regression: Logistic regression is used for binary classification based on statistical methods. It
uses a linear model [14]. Hence, it is used to perform regression on a group of variables [15]. It is a normally
used technique for predicting patterns in data with unambiguous or numeric attributes [14]. It uses a series
of input vectors and a dependent response variable to calculate probability using a logarithm. Probability
lies among the specific class. For binary classification, the response variable is given below:
𝑦𝑖 = {10 (1)

Hence, the formula for calculating that a sample xi belongs in class one is given by
𝑒𝑥𝑝 𝑤 +𝑤 𝑇 𝑥𝑖
𝑃(𝑦𝑖 = 1|𝑥𝑖 ) = 1+𝑒𝑥𝑝(𝑤0 𝑇𝑥 (2)
0 +𝑤 𝑖)

Where W0 and W are the regression standardization parameters, W0 represents the intercept, and W
represents the coefficient vector [16].

4-Random forest: In many research contexts, random forest classification is a well-liked machine learning
technique for creating prediction models. Reducing the number of variables required to produce a forecast
is frequently the aim of prediction modeling, which aims to increase efficiency and lessen the workload
associated with data collecting. There are several variable selection techniques available for random forest
classification settings. However, there isn't much literature to advise users on which technique would be
best for certain dataset types [17].

Fig .2 Representation of random forest

119
 Diaa Salama et al. Journal of Computing and Communication Vol.3 , No.1 , PP. 116-131 , 2024

5-Neural networks: The neural network has good self-learning, self-adapting and generalization ability, but
it may easily get stuck in a local minimum and has a poor convergence rate.[18] As shown in the figure.3,
modeling of input variables as a layer of vertices performed in the network. Then distribution of weight is
applied to every connection within the graph. Moreover, the other vertices are placed into separate levels,
reflecting the distance from the input nodes [19].

Fig .3 a simple neural network [20].

6-Naïve Bayes: is a supervised learning method that doesn't rely on any attribute. The baseline is the Bayes
theorem. Depending on the distribution type. naïve Bayes is a supervised learning method that doesn't rely
on any attribute. The baseline is the Bayes theorem. Based on the kind of distribution, the following
algorithms are available: Three distributions: Bernoulli, Multinomial, and Gaussian. The Bernoulli
distribution is employed in this study to identify fraudulent transactions.[21]
P(x | c)P(c)
𝑃(𝑐|𝑥) = (3)
P(x)
𝑃(𝑐|𝑥) = 𝑃 (𝑥 1 | 𝑐) × 𝑃 (x 2 | 𝑐) ×· · ·× 𝑃 (x n | 𝑐)× 𝑃 (𝑐) (4)
3.3 Performance Metrics
Accuracy, a performance indicator, counts the percentage of examples in a dataset that are properly
classified out of all the instances. F1 score aggregates recall and precision into a single number when there
is an imbalance between the classes in a binary classification problem; it is especially helpful. Recall is a
performance statistic used in classification tasks, sometimes referred to as sensitivity or true positive rate.
Precision is a performance indicator in machine learning and statistics that assesses how well a model makes
good predictions. It is the proportion of correctly predicted true positives to the total of correctly predicted
false positives.

Accuracy = (𝑇𝑁 + 𝑇𝑃)/(𝑇𝑁 + 𝑇𝑃 + 𝐹𝑁 + 𝐹𝑃) (5)

Precision = 𝑇𝑃/𝑇𝑃 + 𝐹𝑃 (6)

Recall = 𝑇𝑃/𝑇𝑃 + 𝐹𝑁 (7)

Specificity = (𝑇𝑁/𝑇𝑁 + 𝐹𝑃) (8)

TP×TN−FP×FN
𝑀𝐶𝐶 = (9)
√(𝑇𝑃+𝐹𝑃)(𝑇𝑃+𝐹𝑁)(𝑇𝑁+𝐹𝑃)(𝑇𝑁+𝐹𝑁)

120
 Diaa Salama et al. Journal of Computing and Communication Vol.3 , No.1 , PP. 116-131 , 2024

4. Experimental Results
The results collected from Gradient Boosting, AdaBoost, Naïve Bayes, Neural network, SVM, CN7 Rule
induction, Logistic Regression, Random Forest, Stochastic gradient Descent, k-nearest Neighbor, Tree,
Constant are shown below.

TABLE II
STATISTICS OF ALGORITHMS WITH 10 K-FOLD

Model AUC CA F1 Perc Recall MCC

AdaBoost 0.749 0.998 0.998 0.998 0.998 0.451
CN7 Rule induction 0.907 0.998 0.998 0.998 0.998 0.486
Constant 0.455 0.998 0.997 0.996 0.998 0.000
Gradient boosting 0.967 0.997 0.997 0.998 0.997 0.368
KNN 0.719 0.998 0.997 0.996 0.998 0.000
Logistic regression 0.929 0.999 0.998 0.999 0.999 0.583
Naive Bayes 0.962 0.997 0.996 0.996 0.997 -0.002
Neural network 0.892 0.998 0.997 0.996 0.998 0.000
Random forest 0.941 0.998 0.998 0.998 0.998 0.421
SVM 0.763 0.998 0.997 0.996 0.998 0.000
Tree 0.455 0.998 0.997 0.996 0.998 0.000

Fig.4 First dataset performance chart with data split

Gradient Boosting leads the pack in machine learning accuracy with an astounding 0.967, closely
followed by Naive Bayes, which performs admirably with 0.962. In contrast, the Constant and Tree
classifiers show the lowest performance, each obtaining a lower accuracy rate of 0.455.

121
 Diaa Salama et al. Journal of Computing and Communication Vol.3 , No.1 , PP. 116-131 , 2024

Table III
STATISTICS OF ALGORITHMS WITH 80/20 DATA SPLIT

Model AUC CA F1 Perc Recall MCC

AdaBoost 0.725 0.998 0.998 0.998 0.998 0.528
CN7 Rule induction 0.911 0.998 0.998 0.998 0.998 0.498
Constant 0.500 0.998 0.997 0.996 0.998 0.000
Gradient boosting 0.820 0.997 0.997 0.997 0.997 0.281
KNN 0.685 0.998 0.997 0.996 0.998 0.000
Logistic regression 0.903 0.998 0.998 0.997 0.998 0.388
Naive Bayes 0.971 0.996 0.996 0.996 0.996 -0.002
Neural network 0.914 0.998 0.997 0.996 0.998 0.000
Random forest 0.948 0.998 0.997 0.998 0.998 0.316
SVM 0.742 0.998 0.997 0.996 0.998 0.000
Tree 0.500 0.998 0.997 0.996 0.998 0.000

Fig. 5 First dataset performance chart with data split

Naive Bayes is the best-performing model, with an accuracy of 0.971; Random Forest comes in second
with 0.948. With a 0.500 accuracy rate, a decision tree is the least accurate model. The following results are
from the second data set.
Table IV
STATISTICS OF ALGORITHMS WITH 10 K-FOLD
Model AUC CA F1 Perc Recall MCC
Tree 0.414 0.993 0.990 0.986 0.993 0.000
SVM 0.573 0.993 0.990 0.986 0.993 0.000
Random forest 0.957 0.995 0.994 0.995 0.995 0.533
Neural Network 0.853 0.993 0.990 0.986 0.993 0.000
Logistic 0,915 0.995 0.994 0.994 0.995 0.523
Regression
Constant 0.414 0.993 0.990 0.986 0.993 0.000
CN2 Rule Induction 0.959 0.992 0.991 0.991 0.992 0.318

122
 Diaa Salama et al. Journal of Computing and Communication Vol.3 , No.1 , PP. 116-131 , 2024

Fig.6. Second dataset performance chart with data split

The highest-performing models are Random Forest (0.957 accuracy) and CN2 Rule Induction (0.959
accuracy), whereas Constant (0.414 accuracy) and Tree (0.414 accuracy) are the lowest-performing models.

Table V
STATISTICS OF ALGORITHMS WITH 80/20 DATA SPLIT

Model AUC CA F1 Perc Recall MCC

Tree 0.500 0.993 0.989 0.985 0.993 0.000
SVM 0.583 0.993 0.989 0.985 0.993 0.000
Random forest 0.943 0.993 0.991 0.990 0.993 0.286
Neural Network 0.851 0.993 0.989 0.985 0.993 0.000
Logistic Regression 0.903 0.995 0.994 0.994 0.995 0.557
Constant 0.500 0.993 0.989 0.985 0.993 0.000
CN2 Rule Induction 0.980 0.993 0.993 0.992 0.993 0.471

Fig.7. Second dataset performance chart with data split

With an astounding accuracy of 0.980, CN2 Rule Induction is the best-performing model; Random
Forest comes in second at 0.943. Constant and Tree models, on the other hand, both score 0.500, which is
the lowest Accuracy.
123
 Diaa Salama et al. Journal of Computing and Communication Vol.3 , No.1 , PP. 116-131 , 2024

Table VI
STATISTICS OF ALGORITHMS WITH 10 K-FOLD

Model AUC CA F1 Perc Recall MCC

AdaBoost 0.854 0.994 0.994 0.995 0.994 0.665
CN7 Rule induction 0.991 0.995 0.995 0.995 0.995 0.663
Constant 0.482 0.992 0.988 0.984 0.992 0.000
Gradient boosting 0.995 0.996 0.996 0.996 0.996 0.767
KNN 0.874 0.993 0.992 0.991 0.993 0.440
Logistic regression 0.944 0.977 0.983 0.993 0.977 0.457
Naive Bayes 0.965 0.985 0.985 0.985 0.985 0.069
Neural network 0.955 0.993 0.990 0.993 0.993 0.388
Random forest 0.991 0.997 0.997 0.997 0.997 0.789
Stochastic gradient Descent 0.515 0.992 0.989 0.992 0.992 0.173
SVM 0.709 0.992 0.988 0.984 0.992 0.000
Tree 0.482 0.992 0.988 0.984 0.992 0.000

Fig.8. Third dataset performance chart with data split

The first three models demonstrate remarkable Accuracy: Gradient Boosting comes in first with 0.995,
closely followed by CN7 Rule Induction and Random Forest, which have excellent Accuracy of 0.991.
Conversely, Constant and Tree, with respective scores of 0.482, share the lowest Accuracy.

124
 Diaa Salama et al. Journal of Computing and Communication Vol.3 , No.1 , PP. 116-131 , 2024

Table VII
STATISTICS OF ALGORITHMS WITH 80/20 DATA SPLIT

Model AUC CA F1 Perc Recall MCC

AdaBoost 0.856 0.995 0.995 0.995 0.995 0.682
CN7 Rule induction 0.988 0.995 0.995 0.995 0.995 0.660
Constant 0.500 0.992 0.988 0.984 0.992 0.000
Gradient boosting 0.992 0.996 0.996 0.996 0.996 0.759
KNN 0.845 0.993 0.992 0.992 0.993 0.451
Logistic regression 0.947 0.978 0.984 0.993 0.978 0.482
Naive Bayes 0.962 0.985 0.985 0.985 0.985 0.084
Neural network 0.969 0.993 0.990 0.993 0.993 0.381
Random forest 0.975 0.997 0.997 0.997 0.997 0.784
Stochastic gradient Descent 0.519 0.992 0.989 0.992 0.992 0.195
SVM 0.720 0.992 0.988 0.984 0.992 0.000
Tree 0.500 0.992 0.988 0.984 0.992 0.000

Fig. 9. Third dataset performance chart with data split

Using gradient boosting, the best-performing model achieves an impressive accuracy of 0.992, with
random forest coming in second with an accuracy of 0. 975. The model's score is 0.500. however, are the
least accurate; these are the constant and three models.

125
 Diaa Salama et al. Journal of Computing and Communication Vol.3 , No.1 , PP. 116-131 , 2024

Table VIII
STATISTICS OF ALGORITHMS WITH 10 K-FOLD
Model AUC CA F1 Perc Recall MCC
Constant 0.414 0.933 0.989 0.985 0.993 0.000
CN2 Rule 0.959 0.992 0.993 0.992 0.993 0.471
Induction
KNN 0.780 0.992 0.989 0.985 0.992 -
0.001
Tree 0.414 0.993 0.989 0.985 0.993 0.000
Random Forest 0.957 0.995 0.992 0.994 0.994 0.446
Gradient 0.993 0.994 0.991 0.990 0.992 0.316
Boosting
SVM 0.582 0.993 0.989 0.985 0.993 0.000
Logistic 0.915 0.995 0.994 0.994 0.995 0.557
Regression
Neural Network 0.853 0.993 0.985 0.985 0.993 0.000

Fig.10. Fourth dataset performance chart with data split

With 0.993 accuracy, Gradient Boosting leads the pack, closely followed by CN2 Rule Induction at
0.959. These top-performing models exhibit outstanding Accuracy. Conversely, at 0.414, Constant and Tree
are the models with the lowest Accuracy.
Table IX
STATISTICS OF ALGORITHMS WITH 80/20 DATA SPLIT

Model AUC CA F1 Prec Recall MCC

Constant 0.500 0.993 0.989 0.985 0.993 0.000
CN2 Rule Induction 0.980 0.993 0.993 0.992 0.993 0.471
kNN 0.679 0.992 0.989 0.985 0.992 -0.001
Tree 0.500 0.993 0.989 0.985 0.993 0.000
Random Forest 0.976 0.994 0.992 0.994 0.994 0.446
Gradient Boosting 0.977 0.992 0.991 0.990 0.992 0.316
SVM 0.612 0.993 0.989 0.985 0.993 0.000
Logistic Regression 0.903 0.995 0.994 0.994 0.995 0.557
Neural Network 0.851 0.993 0.989 0.985 0.993 0.000

126
 Diaa Salama et al. Journal of Computing and Communication Vol.3 , No.1 , PP. 116-131 , 2024

Fig.11. Fourth dataset performance chart with data split

The most accurate model, CN2 Rule Induction, reaches a remarkable accuracy of 0.980. Gradient
Boosting comes in second, with an accuracy of 0.977, very close behind.

Table X
STATISTICS OF ALGORITHMS WITH 10 K-FOLD

Model AUC CA F1 Perc Recall MCC

Constant 0.488 0.993 0.990 0.987 0.993 0.000
CN7 Rule induction 0.985 0.996 0.996 0.996 0.996 0.664
KNN 0.812 0.994 0.992 0.992 0.994 0.358
Tree 0.488 0.993 0.990 0.987 0.993 0.000
Random forest 0.984 0.997 0.997 0.997 0.997 0.776
Gradient boosting 0.997 0.997 0.997 0.997 0.997 0.772
SVM 0.746 0.993 0.990 0.987 0.993 0.000
Logistic regression 0.926 0.979 0.985 0.994 0.979 0.417
Naive Bayes 0.975 0.987 0.988 0.988 0.987 0.105
AdaBoost 0.830 0.996 0.996 0.996 0.996 0.680
Neural network 0.966 0.994 0.992 0.994 0.994 0.364
Stochastic gradient Descent 0.515 0.994 0.990 0.994 0.994 0.171

Fig. 12. Fifth dataset performance chart with data split

127
 Diaa Salama et al. Journal of Computing and Communication Vol.3 , No.1 , PP. 116-131 , 2024

Gradient Boosting (0.997) and CN7 Rule Induction (0.985) are the two best methods in terms of
Accuracy. Random Forest follows closely, obtaining an accuracy of 0.984. However, with respective
accuracy values of 0.488, the Constant and Tree models show the lowest Accuracy.

Table XI
STATISTICS OF ALGORITHMS WITH 80/20 DATA SPLIT

Model AUC CA F1 Perc Recall MCC

Constant 0.500 0.993 0.990 0.986 0.993 0.000
CN7 Rule induction 0.977 0.995 0.995 0.995 0.995 0.650
KNN 0.853 0.994 0.992 0.992 0.994 0.360
Tree 0.500 0.993 0.990 0.986 0.993 0.000
Random forest 0.959 0.997 0.997 0.997 0.997 0.784
Gradient boosting 0.997 0.997 0.997 0.997 0.997 0.797
SVM 0.768 0.993 0.990 0.986 0.993 0.000
Logistic regression 0.890 0.982 0.986 0.993 0.982 0.390
Naive Bayes 0.976 0.988 0.988 0.988 0.988 0.088
AdaBoost 0.835 0.996 0.996 0.996 0.996 0.677
Neural network 0.939 0.994 0.992 0.994 0.994 0.348
Stochastic gradient Descent 0.500 0.993 0.990 0.986 0.993 0.000

Fig. 13. Fifth dataset performance chart with data split

Gradient Boosting and CN7 Rule Induction, the two best performers, both have the highest Accuracy
0.977; Naive Bayes, coming a close second, achieved the second-highest Accuracy of 0.976. Conversely,
the Constant, Tree, and Stochastic Gradient Descent models all had the lowest Accuracy, with a score of
0.500.

128
 Diaa Salama et al. Journal of Computing and Communication Vol.3 , No.1 , PP. 116-131 , 2024

Table XII
STATISTICS OF ALGORITHMS WITH 10 K-FOLD

Model AUC CA F1 perc recall MCC

SVM 0.605 0.993 0.989 0.985 0.993 0.000
Tree 0.500 0.993 0.989 0.985 0.993 0.000
Constant 0.500 0.993 0.989 0.985 0.993 0.000
Naive Bayes 0.936 0.982 0.984 0.987 0.982 0.069
KNN 0.761 0.992 0.989 0.985 0.992 -
0.002
Neural Network 0.894 0.993 0.989 0.985 0.993 0.000
CN2 Rule 0.980 0.993 0.993 0.993 0.993 0.479
induction
Random Forest 0.958 0.995 0.993 0.994 0.995 0.476
Gradient boosting 0.984 0.993 0.993 0.992 0.993 0.441

Fig.14. Sixth dataset performance chart with data split

Using k-fold, Gradient Boosting, and CN2 Rule Induction are the best-performing models, with 0.984
and 0.980 accuracy, respectively. Conversely, Constant is the least accurate model, with an accuracy of
0.500.

Table XIII
STATISTICS OF ALGORITHMS WITH 80/20 DATA SPLIT
Model AUC CA F1 perc recall MCC
SVM 0.575 0.992 0.989 0.985 0.993 0.000
Tree 0.500 0.993 0.989 0.985 0.993 0.000
Constant 0.500 0.993 0.989 0.985 0.993 0.000
Naive Bayes 0.936 0.982 0.984 0.987 0.982 0.069
KNN 0.761 0.992 0.989 0.985 0.992 -0.002
Neural Network 0.894 0.993 0.989 0.985 0.993 0.000
CN2 Rule induction 0.955 0.992 0.992 0.991 0.992 0.400
Random Forest 0.968 0.994 0.993 0.993 0.994 0.477
Gradient Boosting 0.969 0.993 0.992 0.992 0.993 0.420

129
 Diaa Salama et al. Journal of Computing and Communication Vol.3 , No.1 , PP. 116-131 , 2024

Fig.15. Sixth dataset performance chart with data split

With gradient boosting, the best-performing model achieves an Accuracy of 0.969. Random forest
comes in second, with the second-highest Accuracy of 0.968. Conversely, with scores of 0.500, the constant
and Tree models show the least Accuracy.

5. Conclusion
Machine learning is a powerful tool in detecting and preventing online payment fraud. It can analyze
large amounts of data, identify patterns, and make accurate predictions. By leveraging features like
transaction amounts, locations, timestamps, user behavior, and device information, machine-learning
models can identify suspicious patterns and flag fraudulent transactions in real-time. However, they can
produce false positives or negatives, so combining machine learning and human expertise is crucial. The
success of machine learning models depends on the quality and relevance of training data.

130
 Diaa Salama et al. Journal of Computing and Communication Vol.3 , No.1 , PP. 116-131 , 2024

References
[1] 8ir5Sakharova, I. (2012, June). Payment card fraud: Challenges and solutions. In 2012 IEEE international conference on intelligence and security
informatics (pp. 227-234). IEEE
[2] Almazroi, A. A., & Ayub, N. (2023). Online Payment Fraud Detection Model Using Machine Learning Techniques. IEEE Access, 11, 137188-
137203
[3] El Naqa, I., & Murphy, M. J. (2015). What is machine learning? (pp. 3-11). Springer International Publishing.
[4] Minastireanu, E. A., & Mesnita, G. (2019). An Analysis of the Most Used Machine Learning Algorithms for Online Fraud Detection. Informatica
Economica, 23(1).
[5] Nasr, M. H., Farrag, M. H., & Nasr, M. M. (2022). A Proposed Fraud Detection Model based on e-Payments Attributes a Case Study in Egyptian
e-Payment Gateway. International Journal of Advanced Computer Science and Applications, 13(5).
[6] Fang, Y., Zhang, Y., & Huang, C. (2019). Credit Card Fraud Detection Based on Machine Learning. Computers, Materials & Continua, 61(1).
[7] Mijwil, M. M., & Salem, I. E. (2020). Credit card fraud detection in payment using machine learning classifiers. Asian Journal of Computer and
Information Systems (ISSN: 2321–5658), 8(4).
[8] Adepoju, O., Wosowei, J., & Jaiman, H. (2019, October). Comparative evaluation of credit card fraud detection using machine l earning
techniques. In 2019 Global Conference for Advancement in Technology (GCAT) (pp. 1-6). IEEE.
[9] Isabella, S. J., Srinivasan, S., & Suseendran, G. (2020). An efficient study of fraud detection system using Ml techniques. Intelligent Computing
and Innovation on Data Science, 59.
[10] Pumsirirat, A., & Liu, Y. (2018). Credit card fraud detection using deep learning based on auto-encoder and restricted boltzmann
machine. International Journal of advanced computer science and applications, 9(1).
[11] Corballis, M. C., & Nagourney, B. A. (1978). Latency to categorize disoriented alphanumeric characters as letters or digits. Canadian Journal of
Psychology/Revue canadienne de psychologie, 32(3), 186.
[12] Natekin, A., & Knoll, A. (2013). Gradient boosting machines, a tutorial. Frontiers in neurorobotics, 7, 21.
[13] Asim, M., & Zakria, M. (2020). Advanced kNN: A Mature Machine Learning Series. arXiv preprint arXiv:2003.00415.
[14] Ngai, E. W., Hu, Y., Wong, Y. H., Chen, Y., & Sun, X. (2011). The application of data mining techniques in financial fr aud detection: A
classification framework and an academic review of literature. Decision support systems, 50(3), 559-569.
[15] Nusinovici, S., Tham, Y. C., Yan, M. Y. C., Ting, D. S. W., Li, J., Sabanayagam, C., ... & Cheng, C. Y. (2020). Logistic regr ession was as good
as machine learning for predicting major chronic diseases. Journal of clinical epidemiology, 122, 56-69.
[16] Ravisankar, P., Ravi, V., Rao, G. R., & Bose, I. (2011). Detection of financial statement fraud and feature selection using d ata mining
techniques. Decision support systems, 50(2), 491-500.
[17] Speiser, J. L., Miller, M. E., Tooze, J., & Ip, E. (2019). A comparison of random forest variable selection methods for class ification prediction
modeling. Expert systems with applications, 134, 93-101.
[18] Ding, S., Su, C., & Yu, J. (2011). An optimizing BP neural network algorithm based on genetic algorithm. Artificial intelligence review, 36, 153-
162.
[19] Kirkos, E., Spathis, C., & Manolopoulos, Y. (2007). Data mining techniques for the detection of fraudulent fin ancial statements. Expert systems
with applications, 32(4), 995-1003.
[20] West, J., & Bhattacharya, M. (2016). Intelligent financial fraud detection: a comprehensive review. Computers & security, 57, 47-66.
[21] Chen, S., Webb, G. I., Liu, L., & Ma, X. (2020). A novel selective naïve Bayes algorithm. Knowledge-Based Systems, 192, 105361.

131

View publication stats