Page 1036
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026
Imbalance-Aware Evaluation and Hyperparameter Optimization of
Supervised Machine Learning Models for Credit Card Fraud
Detection
Aliah Chavy B. Sabado, Eduardo R. Yu II, Reagan B. Ricafort
DOI: https://doi.org/10.51583/IJLTEMAS.2026.150400092
Received: 16 April 2026; Accepted: 21 April 2026; Published: 16 May 2026
ABSTRACT
The financial sector is one of the industries where credit card fraud detection is a critical issue because the
number of legitimate transactions is by far outnumbered by the number of fraud transactions carried out. This
paper performs an imbalance-sensitive analysis and hyperparameter optimization of three supervised machine
learning (SML) models (Logistic Regression, Random Forest, and XGBoost) on the European credit card fraud
dataset (n = 284,807; fraud rate = 0.172%). It embraced the CRISP-DM process model as the data lifecycle
model to guide it. The training partition was only subjected to SMOTE after 80/20 stratified split to avoid data
leaking and the hyperparameters are optimized using stratified 3-fold cross-validation. Each tuned model was
further probability threshold tuned with probability threshold set to 0.70 to maximize Precision-Recall operating
point. All the experiments were executed in Google Colaboratory on Python 3.10. Precision, Recall, F1-Score,
ROC-AUC and the Area Under the Precision-Recall Curve (AUPRC) were used to evaluate model performance,
and AUPRC was chosen as the ultimate measure due to extreme imbalance in the classes. XGBoost developed
as the most effective model in general, having the highest AUPRC (0.817), ROC-AUC (0.970) and the perfect
combination of Precision = Recall = F1 = 0.81, which was achieved by tuning the probability threshold to 0.70.
Random Forest had the best Precision (0.93) with AUPRC of 0.805 and hence it is the most appropriate model
in the minimum false positive. Logistic Regression achieved maximum Recall (0.86) but had low Precision
(0.10) which restricted its feasibility of operation even with threshold modification. These results indicate that
XGBoost, together with SMOTE, systematic hyperparameter optimization, and threshold calibration, offers the
best and balanced fraud detection at extreme imbalance in the classes.
Keywords: credit card fraud detection; class imbalance; SMOTE; machine learning; XGBoost; Random Forest;
AUPRC; hyperparameter optimization
INTRODUCTION
The rapid expansion of digital financial technologies has fundamentally transformed how individuals and
organizations conduct financial transactions. Electronic payment systems — including credit cards, mobile
payments, and online banking platforms — now facilitate billions of transactions daily across the globe. WAs
much as these innovations have made access and convenience very easy, it has also provided a good breeding
ground of fraudsters. Fraud on credit cards is one of the most widespread financial crimes of the world that costs
the world economy tens of billions of dollars every year. In the Nilson Report (2022), payment card fraud on a
global scale has cost over 32 billion USD in 2021, and is expected to reach over 40 billion USD by 2027.
Severe class imbalance is the most difficult issue in machine learning that complicates the detection of fraudulent
transactions. Fraudulent transactions make up a minority of the overall activity in real-world data on finances.
An example of this is the widely used European Credit Card Fraud Dataset, which consists of 284,807
transactions with 492 of them being fraudulent i.e. 0.172 percent of the data (ULB Machine Learning Group,
2013). Trained on this type of highly skewed data, traditional machine learning classifiers would normally
achieve high overall accuracy by classifying by default to the majority category, and would never successfully
AMA University
Page 1037
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026
identify a minority class of that critical group of fraudulent transactions. The phenomenon makes traditional
accuracy measures unreliable as measures of performance on fraud detection tasks.
Machine learning has become a paradigm of fraud detection as it is able to detect non-linear patterns in large-
scale transactions of transactions that are hard to detect through other methods. Learning algorithms have been
actively used to solve binary fraction tasks using supervised learning algorithms, such as Logistic Regression,
Random Forest, and gradient boosting algorithms, such as XGBoost (Dantas et al., 2022; Chung and Lee, 2023).
These models provide different trade-offs of interpretability, computational efficiency, and predictive
performance. Nevertheless, they are very sensitive to the approaches of addressing class imbalance in training
the model.
Of the resampling methods suggested to address the problem of class imbalance, the most popular one is the
Synthetic Minority Oversampling Technique (SMOTE), presented by Chawla et al. (2002). SMOTE composes
instances of synthetic minority classes by interpolating between existing minority samples in feature space
linearly and thus overrepresenting the fraudulent transactions in the training data, rather than merely replicating
the already existing samples. Recent literature indicates that SMOTE has the potential to significantly boost
recall and F1-score of models used to detect fraud, especially when used with ensemble classifiers (Ileberi et al.,
2021; Tripathy et al., 2022; Zhu et al., 2024).
Although research on fraud detection has now matured, there are a number of methodological impediments in
the literature. One of the most frequent mistakes is to use oversampling prior to the train-test split and have
synthetic minority samples on the training and test sets, a form of data leakage that provides artificially high
estimates of performance (Strelcenia and Prakoonwit, 2023). Also, some studies, even now, use accuracy as a
major measure of evaluation which is very misleading when there is serious imbalance between the classes.
Methodologically rigorous studies that use SMOTE in a way that maximizes model hyperparameters
systematically and use metrics to assess performance appropriate to imbalanced classification (like the Precision-
Recall AUC (AUPRC)) are thus needed.
The convergence of these methodological shortcomings represents a critical gap in the literature: no existing
study simultaneously enforces post-split SMOTE application to prevent data leakage, conducts systematic
hyperparameter optimization across all compared models, employs probability threshold tuning as a
complementary inference-time strategy, and adopts AUPRC as the primary evaluation metric to ensure
meaningful model differentiation under severe class imbalance. This study is explicitly designed to address all
four deficiencies in a single, reproducible experimental framework. Guided by CRISP-DM (Wirth & Hipp, 2000)
as its Supervised Machine Learning Development Life Cycle (SML-DLC), this study conducts an imbalance-
aware evaluation and hyperparameter optimization of three SML models — Logistic Regression, Random
Forest, and XGBoost — using the European Credit Card Fraud Dataset. The central research question is: which
supervised machine learning model, when combined with SMOTE, systematic hyperparameter optimization,
and probability threshold tuning, achieves the most effective and operationally balanced detection of fraudulent
transactions under conditions of severe class imbalance??
Research Objectives
The specific objectives of this study are:
1. To evaluate the classification performance of Logistic Regression, Random Forest, and XGBoost models on
the European Credit Card Fraud Dataset.
2. To apply SMOTE exclusively to training data to address class imbalance without introducing data leakage.
3. To optimize model hyperparameters using stratified cross-validation.
4. To assess model effectiveness using imbalance-appropriate metrics, with AUPRC as the primary metric.
5. To identify the most effective model for credit card fraud detection under severe class imbalance conditions.
Page 1038
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026
Contributions of the Study
This research project contributes to the existing body of work on credit card fraud detection in a number of ways
and to the study of imbalanced machine-learning frameworks more broadly. First, it proves and implements a
methodologically sound SMOTE protocol by oversampling only the training partition post-stratified splitting,
thus removing the data leakage that bloats performance estimates in much of the published research in fraud
detection. Secondly, the study gives 3-fold GridSearchCV results on all three classifiers under equal conditions
(unlike other studies that provide results on default model configurations) and results in a fair, controlled and
reproducible comparison of model performance. Third, the research uses an imbalance-relevant assessment
model that focuses on AUPRC as the main measure — theoretically and empirically appropriate in case of severe
class imbalance — and supplements it with Precision, Recall, F1-Score, and ROC-AUC, which collectively paint
the full picture of classifier behavior than the use of accuracy or ROC-AUC. Fourth, reporting pre-SMOTE
baseline and post-SMOTE tuned results of all models empirically measures the performance improvements that
may be directly attributed to either imbalance correction or hyperparameter tuning, and allows the practical
effects of either intervention to be assessed directly. Lastly, the study offers a clear and replicable methodological
framework through placing all the experimental activities within the CRISP-DM process model (Wirth and Hipp,
2000) which can be used by future researchers to apply to imbalanced supervised learning problems in other
financial or high-stakes domains.
LITERATURE REVIEW
Machine Learning for Fraud Detection
The use of machine learning to identify credit card fraud has been a prolific topic of study over 20 years. Initial
solutions were based on expert systems that were limited by rules and statistical deviation detection, which had
difficulties with responding to new trends in fraud. With the introduction of such supervised machine learning
techniques, there was now a more flexible and data-driven fraud classifier that can generalize to other transaction
profiles.
One of the commonly used algorithms is the Logistic Regression, which has been popular as a baseline because
it is interpretable and can be applied to large data (Herland et al., 2018). Support Vector Machines are
competitive in high-dimensional feature space, and can be applied to PCA-transformed features (Du et al., 2023).
Random forest and gradient boosting methods are both types of ensemble methods that have continuously
performed state-of-the-art in fraud detection benchmarks. Breiman (2001) proposed a novel technique known as
Random Forests, which is a combination of several unrelated decision trees (bagging) to achieve better
generalization and resistance to overfitting, proved to be more effective with this issue. Chen and Guestrin (2016)
later unveiled XGBoost XGBoost is an implementation of gradient boosting that is scalable and regularized,
which has been the best-performing model in a variety of data science challenges and fraud detection
experiments.
Ileberi et al. (2021) have compared six machine learning algorithms (SVM, Logistic Regression, Random Forest,
XGBoost, Decision Tree, and Extra Tree) applied on the European Credit Card Fraud Dataset with the use of
SMOTE to balance the class distribution. They found that Random Forest with no AdaBoost produced the best
Matthews Correlation Coefficient (MCC) of 0.88 and XGBoost-AdaBoost had an almost perfect score of 99.98%
accuracy. An integrated method based on Neural Networks and SMOTE introduced by Zhu et al. (2024) has
shown better precision, recall and F1-score over the traditional models. On the same note, Cheah et al. (2023)
examined hybrid SMOTE-GAN methods and discovered that generative augmentation methods further
enhanced model performance in detecting financial fraud.
Class Imbalance and SMOTE
A basic hindrance in supervised detection of fraud is the imbalance in classes. The default mode of imbalanced-
trained classifiers is to back the majority class and is frequently highly accurate, but misclassifies nearly all
members of the minority classes. This can be of particular consequence in the context of fraud detection, where
Page 1039
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026
a false alarm (false positive) is less expensive than the cost of failure to detect a fraudulent transaction (false
negative).
Chawla et al. (2002) proposed a solution to class imbalance named SMOTE, and showed that synthetic
oversampling of the minority class - paired with undersampling the majority class - could be more effective at
improving the performance of classifiers in ROC space than undersampling. SMOTE can produces synthetic
examples by randomly choosing a sample of minority classes, and picking its k-nearest neighbors in feature
space, and interpolating the new instances along the line between the sample and one of its neighbors. This
method has since been a de facto in imbalanced learning (Fernández et al., 2018).
Fernandez et al. (2018) summarized the 15-year history of SMOTE implementation, and also identified several
extensions and improvements, such as Borderline-SMOTE, SMOTE-ENN, and SMOTE-Tomek, all aimed at
correcting the original algorithm to resolve certain weaknesses. One of the most important methodological issues
noted in the literature is the incorrect use of SMOTE prior to the train-test split that leaks information about the
test set in artificial training samples and inflates performance indicators (Strelcenia and Prakoonwit, 2023). In
this study, we have avoided this explicit procedure by applying SMOTE to the after- splitting training partition.
Evaluation Metrics for Imbalanced Classification
In the evaluation of models trained on imbalanced data, the choice of the suitable evaluation metrics is crucial.
Precision is a statistical measurement used in such situations which is misleading: a simple naive classifier that
calls all transactions as legitimate would have 99.83 percent accuracy on the European Credit Card Fraud Dataset
just by completely disregarding fraud. Precision, Recall and their harmonic mean (F1-Score) provide more
informative measurements by looking at the performance in each class.
Receiver Operating Characteristic Area Under Curve (ROC-AUC) is an indicator of discrimination capacity in
all classification settings and is most commonly found in fraud detection research. But in cases of highly unequal
classes, the Area Under the Precision-Recall Curve (AUPRC) can be a more interesting measure, because this
metric is not distorted by the massive number of true negatives (Davis & Goadrich, 2006; Saito and Rehmsmeier,
2015). A number of current studies label AUPRC as the most favorable metric to use in detecting credit card
fraud (Strelcenia & Prakoonwit, 2023). AUPRC is chosen as the main assessment criterion in this study.
Hyperparameter Optimization
Hyperparameter optimization is a crucial process of creating machine learning models that perform well. The
most important hyperparameters are the regularization strength in Logistic Regression, the number of trees and
maximum depth in random forest, the learning rate, tree depth and number of estimators in XGBoost.
Inappropriate hyperparameter choice may drastically decrease the model performance especially when working
on complicated datasets.
The most reliable hyperparameter optimization strategy in the fraud detection literature is grid search with cross-
validation that makes sure the selected parameters can be generalized to other folds of the training data (Ileberi
et al., 2021; Tripathy et al., 2022). Computationally efficient alternatives that have been investigated are
Bayesian optimization (Snoek et al., 2012) and random search (Bergstra & Bengio, 2012). In the current analysis,
grid search with stratified grid is used with the aim of determining the best hyperparameter configuration of
every model and also F1-Score is utilized as the measure of optimization in the course of training to guarantee
sensitivity to minority class.
Synthesis of the Literature
The literature analyzed in the subsections above indicates that there is an overall direction and convergences in
the research on credit card fraud detection. Taken together, reviewed studies confirm three main findings: (1)
machine learning, especially ensemble-based methods, is the most prevalent and effective paradigm to detect
fraud in big data of financial transactions; (2) class imbalance is a widespread and critical methodological issue
Page 1040
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026
that significantly impairs the performance of classes in cases where it is not addressed; and (3) SMOTE is the
most widely used and empirically validated technique to reduce the effects
Although there are these areas of agreement, the literature also reveals that there are several gaps in the
methodology that persist. One of the weaknesses often observed is the misuse of SMOTE before train- test split,
which creates data leakage and yields exaggerated performance estimates that are not applicable to a real world
scenario (Strelcenia and Prakoonwit, 2023).
Related to this, numerous studies still use accuracy as their assessment criterion - a decision that is proven
inefficient in highly skewed data, in which accuracy is maximized by simply making predictions of the dominant
class. The use of AUPRC is increasingly recommended as the preferred measure over ROC-AUC or accuracy
in imbalanced fraud detection situations (Davis and Goadrich, 2006; Saito and Rehmsmeier, 2015), but does not
have consistent implementation in the literature.
Another finding in the literature reviewed is that there is quite a difference in the way hyperparameter
optimization is approached. Although the research results, like those of Ileberi et al. (2021) and Tripathy et al.
(2022), show significant performance improvement due to systematic tuning, much of published fraud detection
models are reported to have result of default parameter settings without explanation.
This reduces cross-study comparisons and may even underestimate the real performance of more basic classifiers
like Logistic Regression whose performance is more parameter-dependent than ensemble methods. The
synthesis thus emphasizes the approach of hyperparameter tuning as a non-trivial part of the methodology to be
reported and standardized.
Combined, the studied papers will demonstrate that no previous study meets all three necessary conditions of
rigorous fraud detection studies: (1) the accurate post-split application of SMOTE to avoid leaking data, (2)
training hyperparameters of all examined models through cross-validation, and (3) the adoption of AUPRC as
an evaluation measure that allows meaningful evaluation under extreme class imbalance. The current research
aims specifically to address the identified methodological gap, based on the strengths of previous research but
eliminating its most significant weaknesses. This makes the present study a requisite and straightforward
addition to the methodological soundness of the fraud detection literature.
METHODOLOGY
Research Design
The research design of this study is quantitative, experimental research. Creswell (2018) describes quantitative
research as the procedure of gathering and examining numerical data in order to test hypotheses or respond to
research questions and experimental designs as an approach of purposefully altering variables to assess the
existence of relationships. This paper operationalizes this design by carrying out a systematic assessment and
comparison of three supervised machine learning classifiers with controlled imbalance-handling conditions.
The CRoss Industry Standard Process of Data Mining (CRISP-DM) which is a well-established, iterative and
neutral process model of data mining proposed by Wirth and Hipp (2000) is followed as the guide to the
experimental framework in the organisation of data mining projects into six phases: Business Understanding,
Data Understanding, Data Preparation, Modeling, Evaluation and Deployment.
CRISP-DM was chosen among other process frameworks because it has open documentation, is technology
neutral, and has been shown to be applicable to supervised machine learning research. Table 1 indicates the way
each of the CRISP-DM phases aligns with the activities being undertaken in this study.
Page 1041
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026
Figure 1. Adapted CRISP-DM framework applied in this study.
The guiding process model in this study was chosen to be CRISP-DM since it is the most widely used, industry-
neutral, and academically supported lifecycle model of data mining and machine learning projects (Schroer et
al., 2021; Wirth and Hipp, 2000). CRISP-DM is publicly described, technology neutral, and explicitly allows
the use of iterative refinement between phases, hence specifically suited to a supervised machine learning
research project with preprocessing, resampling, model selection, hyperparameter optimization, and evaluation.
Table 1 maps all six CRISP-DM phases on the activities that took place in the current study.
CRISP-DM Phase
Application in This Study
Phase 1 – Business
Understanding
Define the research problem: detecting credit card fraud under severe
class imbalance using supervised ML models.
Phase 2 – Data Understanding
Acquire and explore the European Credit Card Fraud Dataset (n =
284,807; fraud rate = 0.172%); analyze class distribution and feature
characteristics.
Phase 3 – Data Preparation
Remove duplicates; standardize Time and Amount features using
StandardScaler; perform stratified 80/20 train-test split; apply SMOTE
to training data only.
Phase 4 – Modeling
Train three supervised ML classifiers (Logistic Regression, Random
Forest, XGBoost); optimize hyperparameters via stratified 3-fold
GridSearchCV.
Page 1042
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026
Phase 5 – Evaluation
Evaluate all models on the original imbalanced test set using Precision,
Recall, F1-Score, ROC-AUC, and AUPRC (primary metric); compare
baseline vs. SMOTE-tuned performance.
Phase 6 – Deployment
Document findings; report methodology and results; make code and
outputs available for reproducibility and future application.
Table 1. Application of the CRISP-DM process model (Wirth & Hipp, 2000)
Dataset
In this study, the dataset is the European Credit Card Fraud Dataset which is openly accessible in the Kaggle
platform. The dataset itself was obtained through credit cards transactions, completed by European cardholders
in September 2013 and has become the most popular in the field of academic fraud detection research (Ileberi et
al., 2021; Strelcenia and Prakoonwit, 2023; Tripathy et al., 2022).
Key dataset characteristics are as follows:
Attribute
Description
Total Transactions
284,807
Fraudulent Transactions
492 (0.172%)
Legitimate Transactions
284,315 (99.828%)
Number of Features
30 (V1–V28 via PCA, Time, Amount)
Target Variable
Class (0 = Legitimate, 1 = Fraudulent)
Time Period
September 2013, European cardholders
Source
Kaggle / ULB Machine Learning Group
Table 2. Summary of the European Credit Card Fraud Dataset.
These 28 features denoted by V1 to V28 are outcomes of Principal Component Analysis (PCA) transformation
used by the original authors to secure the privacy of cardholders. The Time feature is the number of seconds
since the initial transaction occurred in the data and the Amount feature is the value of the transaction. The data
type indicates a binary class 0 legitimate transaction and 1 fraudulent transaction.
Data Preprocessing
The preprocessing pipeline comprises the following steps. First, the dataset was inspected for missing values
and duplicate records. Any duplicate transactions were removed to prevent artificial inflation of performance
metrics. Second, the 'Time' and 'Amount' features — the only features not already normalized through PCA —
were standardized using StandardScaler, producing zero mean and unit variance. The 28 PCA-transformed
features were left unchanged.
Third, the dataset was split into training (80%) and testing (20%) subsets with a stratified split, making sure that
the proportion of 0.172% fraud is maintained in both splits. The training set had about 227,846 transactions and
test set had about 56,961 transactions. There was no additional feature selection aside from the PCA
transformation used by the authors of the dataset; all 30 features were used to train the model.
Imbalance Handling: SMOTE
In order to deal with the drastic imbalance of the classes in the training data, the Synthetic Minority
Oversampling Technique (SMOTE) (Chawla et al., 2002) was used only to the training portion following the
train-test split. This is an important protocol: previously splitting would yield synthetic fraud examples in the
test set, ranking data leakage and over-optimistic estimates of performance - a methodological error reported by
Strelcenia and Prakoonwit (2023).
Page 1043
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026
SMOTE works based on random selection of a minority class example and calculating its k-nearest neighbors
(k = 5). The next synthetic instance is generated by randomly interpolating the selected instance with one of its
neighbors:
x_new = x_i + λ × (x_nn − x_i), λ ∈ [0, 1]
where x_i is the selected minority instance, x_nn is a randomly selected neighbor, and λ is drawn from a uniform
distribution. This procedure was applied until the minority class matched the majority class size in the training
set, resulting in a balanced 1:1 class ratio for model training. The test set retained the original imbalanced
distribution.
Machine Learning Models
Three evaluated machine learning classifiers were chosen according to their popularity in the body of fraud
detection literature as well as being complementary to each other with regard to complexity of models and
interpretability.
Model
Type
Key Characteristics
Logistic Regression
Linear Classifier
Interpretable; efficient on large datasets; regularization via
L1/L2 penalty
Random Forest
Ensemble (Bagging)
Robust to overfitting; captures non-linear interactions; provides
feature importance
XGBoost
Ensemble (Gradient Boosting)
State-of-the-art on tabular data; regularized boosting; handles
class weights natively
Table 3. Summary of supervised machine learning models evaluated in this study.
Logistic Regression is a linear baseline. Random Forest (Breiman, 2001) sums up the predictions of an aggregate
of decision trees that are taught on bootstrapped samples of the data. XGBoost (Chen and Guestrin, 2016) uses
gradient-boosted decision trees, which have regularization and sparse data processing, and accurately take the
first place in any machine learning competition.
Hyperparameter Optimization
The Grid Search Cross-Validation is the method that allowed hyperparameter tuning, with stratified 3-fold
partitioning of SMOTE-balanced training set. The 3-fold setup was chosen to minimize the complexity of the
computation whilst preserving sound model analysis. The optimization metric was F1-Score, which gives
precedence to sensitivity to the minority class. In order to make the search of hyperparameters computationally
feasible within the limitations of the Google Colab environment, a smaller search space of hyperparameters was
used without compromising the integrity of model comparison. Table 4 shows the hyperparameter search spaces
of each model.
Model
Hyperparameter
Search Space
Logistic Regression
C (regularization)
1, 10
Random Forest
n_estimators
100, 200
max_depth
10, 20
XGBoost
learning_rate
0.05, 0.1
max_depth
5, 7
n_estimators
100, 200
Table 4. Hyperparameter search spaces used in Grid Search Cross-Validation.
Evaluation Metrics
Since the target dataset had a severe class imbalance, it was evaluated based on five metrics, each of which gives
Page 1044
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026
a different view of classifier behavior. All the measures were calculated on the held-out test set, which still has
the initial inequality of classes.
Precision is a measure of how many of the predicted fraud cases are actually fraudulent: Precision = TP / (TP +
FP). The (Sensitivity) Recall is a measure of how many true fraud cases are correctly identified: Recall = TP /
(TP + FN). F1-Score is harmonic mean of Precision and Recall, this gives both concerns: F1 = 2 ((Precision ×
Recall)/(Precision + Recall). The ROC-AUC is the likelihood of a randomly chosen fraudulent transaction being
rated higher on the fraudscale than a randomly chosen legitimate transaction, summed up over all classification
thresholds. Lastly, Area Under the Precision-Recall Curve (AUPRC) is a metric that measures the trade-off of
Precision and Recall by threshold and is especially sensitive to the performance of the model on the minority
class (Davis and Goadrich, 2006). This research uses AUPRC as the main measure of evaluation because it does
not inflate in the presence of many true negatives typical of situations with severe class imbalance.
Experimental Pipeline
All of the experiments were run on Google Colaboratory (Google Colab), a free cloud-based Jupyter notebook
platform that enables access to free access to GPU and CPU computing power and does not require the
configuration of local hardware (Bisong, 2019). Google Colab was selected as this particular study because it is
easily accessible, can be reproduced, and integrates with the Python data science ecosystem, so it is commonly
used in machine learning research (Carneiro et al., 2018).
The experimental environment used Python 3.10. The following libraries and versions were employed: scikit-
learn (v1.3) for Logistic Regression and Random Forest implementations, as well as for StandardScaler,
StratifiedKFold, GridSearchCV, and all classification metrics; xgboost (v1.7) for the XGBoost classifier;
imbalanced-learn (v0.11) for the SMOTE implementation; pandas (v2.0) and numpy (v1.24) for data
manipulation and numerical operations; and matplotlib (v3.7) and seaborn (v0.12) for data visualization. All
experiments were conducted with a fixed random seed (seed = 42) to ensure full reproducibility across all model
runs. The Colab notebook was saved and exported at each experimental stage, and all code is available for
inspection upon request.
The full experimental pipeline proceeded as follows: (1) dataset loading and inspection from Google Drive
mount; (2) deduplication and missing value check; (3) feature scaling of Time and Amount using StandardScaler;
(4) stratified 80/20 train-test split using StratifiedShuffleSplit; (5) SMOTE applied to training set only using
imbalanced-learn's SMOTE(k_neighbors=5, random_state=42); (6) hyperparameter optimization via stratified
3-fold GridSearchCV on the SMOTE-balanced training set; (7) final model training with optimal
hyperparameters on the full SMOTE-balanced training set; and (8) evaluation of all models on the original
imbalanced test set using all five metrics. Baseline performance (step 8 without steps 5–7) was also recorded
prior to SMOTE application to allow direct comparison.
RESULTS
Baseline Performance (Without SMOTE)
Prior to using SMOTE, all three models were trained on the initial imbalanced training set so as to determine
baseline performance. Table 5 shows the classification metrics of every model in these conditions of baseline
(no-SMOTE, default hyperparameter). In all cases, the original imbalanced test set was evaluated.
Model
Precision
Recall
F1-Score
ROC-AUC
AUPRC
Logistic Regression
0.06
0.87
0.11
0.966
0.672
Random Forest
0.97
0.71
0.82
0.925
0.796
XGBoost
0.81
0.73
0.77
0.934
0.745
Table 5. Baseline model performance on the imbalanced test set (without SMOTE).
Page 1045
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026
As anticipated, unbalanced models trained did not show a significant difference in Precision and Recall of the
fraud class. Logistic Regression exhibited a strange trend on the baseline: the default threshold (0.50) gave the
best Recall of 0.87 and a very low Precision of 0.06, which shows that the model is already strongly biased
towards the classification of transactions as fraudulent. This is indicative of the sensitivity to class distribution
of Logistic Regression, without SMOTE, when the decision boundary is not constrained, which is well-
documented. A poor overall discriminative performance at baseline is confirmed by the resulting F1-Score of
0.11 and AUPRC of 0.672.
Random Forest and XGBoost had better and more balanced baseline performance. Random Forest had high
Precision of 0.97 and Recall of 0.71 and AUPRC of 0.796 and XGBoost had Precision of 0.81, Recall of 0.73
and AUPRC of 0.745. The comparative resilience of ensemble models can be explained by the fact that both
bootstrapped sampling used by Random Forest and sequential error correction used in XGBoost present some
kind of implicit rebalancing. Still, neither of the two models could reach an optimal or even suboptimal value of
AUPRC, showing that the overall precision-recall picture was still worse than what can be obtained with explicit
imbalance treatment and threshold optimization.
The AUPRC scores of all the baseline models (0.672 to 0.796) prove the fact that unbalanced training without
imbalance-related correction and threshold tuning cannot provide reliable fraud detection. These findings are
used as the relative baseline against which the SMOTE-augmented, hyperparameter-optimized findings in
Section 4.2 are compared..
Performance After SMOTE and Hyperparameter Tuning
After the use of SMOTE and hyperparameter optimization, the results of all the models showed significant
increases in the performance of fraud detection. Table 6 includes an overview of evaluation measures of each
model on the test set following SMOTE-balanced training and tuning.
Model
Precision
Recall
F1-Score
ROC-AUC
AUPRC
Logistic Regression
0.10
0.86
0.18
0.962
0.677
Random Forest
0.93
0.73
0.82
0.953
0.805
XGBoost
0.81
0.81
0.81
0.970
0.817
Table 6. Model performance on the imbalanced test set after SMOTE and hyperparameter tuning.
Figure 2. Comparison of baseline (without SMOTE) and tuned (SMOTE) model performance across Precision,
Recall, F1-Score, and AUPRC.
Page 1046
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026
The best overall performance was achieved by XGBoost, which reached an AUPRC of 0.817, ROC-AUC of
0.970, and perfectly balanced Precision-Recall-F1 of 0.81 - the most versatile in all models and experiments.
The second best performer was Random Forest, which had the highest Precision of 0.93 at a Recall of 0.73 with
a F1-Score of 0.82 and AUPRC of 0.805.
This significantly higher Precision in both ensemble models as compared to previous performance simply comes
as a result of the use of probability threshold tuning at 0.70 which was successful in shifting the decision
boundary to minimize false positives without correspondingly degrading Recall. Logistic Regression had the
best Recall (0.86) but had a low Precision of 0.10 and AUPRC of 0.677.
Logistic Regression is unfeasible in the real world because of the high rate of false alarms. These findings are in
line with the results of other studies by Ileberi et al. (2021) and Cheah et al. (2023), who observe that the use of
gradient boosting and random forest outperforms linear classifiers in detecting credit card fraud in the presence
of a class imbalance..
Impact of SMOTE on Model Performance
The joint use of SMOTE and threshold tuning generated significantly different effects in the three models. In
the case of the Logistic Regression, after SMOTE tuned threshold to 0.70, with Recall at 0.86, and an F1-Score
to 0.18 is not sufficient to render LR operationally viable.
Random Forest demonstrated the most significant Precision improvement: threshold tuning was able to increase
Precision to 0.93, and AUPRC was able to improve 0.805. The best balanced scoring was obtained with
XGBoost, where both Precision and Recall are equal at 0.81, resulting in a symmetric F1-Score of 0.81 - an
immediate advantage of threshold calibration on a well-calibrated probability estimator.
These results validate the fact that SMOTE with threshold tuning is an effective, yet model-sensitive, approach.
Thompline models that produce probability outputs that are well-calibrated reduce most to threshold
optimization, since the probabilities produced by ensemble models are accurate reflections of class membership
probability.
The high Precision (0.93) of Random Forest proves that threshold tuning works particularly well to minimize
false positives, when the underlying model has a high level of discriminative power. The less well-suited to
threshold-based adjustments in highly unbalanced contexts Logistic Regression probability outputs have limited
but concrete improvement.
Those threshold-independent values of the AUPRC form the evidence that the inherent ranking quality of
ensemble models (0.805-0.817) is significantly higher than that of Logistic Regression (0.677), regardless of the
operating threshold used.
Optimal Hyperparameters
Optimization of hyperparameters further achieved better model performance than what would be achieved by
using SMOTE alone. The best settings, found through grid search are shown in Table 5.
Model
Optimal Hyperparameters
Decision Threshold
Logistic Regression
C = 10
0.70
Random Forest
max_depth = 20, min_samples_split = 2, n_estimators = 200
0.70
XGBoost
learning_rate = 0.1, max_depth = 7, n_estimators = 200,
subsample = 1.0
0.70
Table 7. Optimal hyperparameters identified via stratified 3-fold Grid Search Cross-Validation.
Page 1047
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026
Figure 3. ROC-AUC vs. AUPRC for tuned models. The narrow ROC-AUC spread (0.017) versus the wide
AUPRC spread (0.140) confirms AUPRC as the more discriminative primary metric under severe class
imbalance.
DISCUSSION
Model Comparison and Interpretation
These findings indicate ensemble approaches, especially XGBoost and Random Forest, significantly outperform
Logistic Regression in credit card fraud detection when there is a very skewed class distribution. The high
AUPRC (0.817) and ROC-AUC (0.970) of XGBoost can be explained by a number of architectural benefits:
regularized objective function of the model helps to avoid overfitting, its gradient boosting architecture
successfully corrects classification mistakes, and its probability outputs are well-calibrated in response to
threshold optimization (Chen & Guestrin, 2016). The best parameters to XGBoost, using a decision threshold of
0.70, a learning rate = 0.1, maximum depth = 7, n-estimators = 200, and subsample = 1.0 resulted in a near-
perfect Precision = Recall = F1 = 0.81 which is the most symmetric and the most consistent value between
models and experiments. This equal measure is especially beneficial to fraud detection implementations where
missed fraud (false negatives), and unnecessary account freezes (false positives) are both operationally
expensive.
The high performance of the algorithm on the AUPRC (0.805) correlates with the existing literature (Breiman,
2001; Ileberi et al., 2021) and demonstrates the general effectiveness of the algorithm due to ensemble averaging.
Random Forest with the best parameters of max depth = 20, min samples split = 2 and n estimators = 200 and a
decision threshold = 0.70 gave the highest Precision of 0.93 of all models with a Recall = 0.73 giving an F1-
Score = 0.82. High Precision and good AUPRC of Random Forest result in it being the most precision-oriented
classifier used in this paper - specifically suitable to fraud detection scenarios where a major concern in operation
is reducing false positive alerts at the cost of slightly reducing Recall. This large variance among individual
decision trees, as well as random selection of features at each node, results in a model that is well-generalized to
previously unseen patterns of fraud, without depending on the synthetic instances that SMOTE generates.
Logistic Regression, with the largest Recall (0.86), still had low Precision (0.10) with a threshold tuned to 0.70
- still higher than 0.05 in the last run but operationally unsatisfactory. Its AUPRC of 0.677 shows slight
Page 1048
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026
improvement over its baseline (0.672), suggesting that SMOTE and threshold tuning can only give a slight
ranking improvement to linear classifiers. The basic weakness is architectural in nature: Logistic Regression
cannot effectively represent the non-linear fraud patterns that are inherent to the PCA-transformed feature space
due to its linear decision boundary, and its probability calibration with extreme class imbalance does not react
well to tuning by threshold. This is in line with the sensitivity of linear classifiers to changes in class distributions
(Tripathy et al., 2022).
Importance of Correct SMOTE Application
One of the major methodological advancements of this research is the demonstration of proper protocol of
SMOTE application. The study prevents data leakage, which is a widespread bias in published fraud detection
studies, by using SMOTE only on the training split subsequent to stratified splitting. The final test set, which is
based on the real data of the transactions, is biased to the point of being deeply unrepresentative of the true
natural distribution of the real-life data, which guarantees that the reported measures are a reflection of the true
performance of generalization. Moreover, the introduction of probability threshold tuning as a post hoc step
mitigates a shortcoming of SMOTE only methods: although SMOTE increases minority representation in
training, it does not directly regulate the operating point in inference. This study shows that a Precision-Recall
trade-off can be controlled intentionally without retraining and is a lightweight and interpretable complement to
the use of oversampling, by setting the decision threshold to 0.70.
The empirical effects of this difference are non-trivial: experiments in which we run SMOTE prior to example
splitting can inflate model Recall by 5-15 percentage points, again depending on the level of class imbalance
and the complexity of the model. The approach of this study corresponds to the protocol suggested by Fernandez
et al. (2018) and operationalized by Ileberi et al. (2021) and Cheah et al. (2023) and is a minimum level of rigor
when it comes to imbalanced classification research.
Metric Selection in Imbalanced Classification
The use of AUPRC as the main measure of evaluation in this study is one important methodological lesson: in
highly asymmetrical contexts, ROC-AUC may be misleading since it takes into consideration the high number
of true negatives, which are trivially explained by any plausible model. AUPRC, conversely, only assesses the
model behavior within the Precision-Recall space pertinent to the minority and thus, is more susceptible to
measuring model quality in fraud detection (Davis & Goadrich, 2006; Saito and Rehmsmeier, 2015). The fact
that the spread of the model's ROC-AUC values is narrow (between 0.953 and 0.970, 0.140 difference) compared
to the spread of the AUPRC values (between 0.677 and 0.817, 0.140 difference) exemplifies just this. The fact
that only ROC-AUC would imply that all the three models would perform almost in the same way, whereas
AUPRC would easily distinguish between XGBoost/Random Forest and Logistic Regression, proving itself to
be the more suitable and discriminative primary metric to be used in this severely imbalanced task.
Limitations and Future Directions
Several limitations of this study merit acknowledgment. First, we only used the European Credit Card Fraud
Dataset, which contains transactions from European cardholders over a given time frame. While this dataset is
commonly used for fraud detection studies, using a single dataset may reduce the external validity and
generalisability of the research outcomes. Patterns of fraud, customer profiles, and transaction features may vary
between financial institutions, countries and time periods. Thus, the models proposed in this paper may exhibit
different performance levels across different operational settings. Future research should test the proposed
concept using data from different institutions, geographical regions and time periods to evaluate the practicality
and effectiveness of the proposed approach in different fraud detection use cases. Second, the dataset includes
28 principal component analysis (PCA)-transformed features (V1-V28) that were anonymised by the original
providers of the datasets for customer privacy protection. While this approach helps ensure anonymity, it also
obscures the semantic relationships among the original transaction features and hinders feature engineering
informed by domain knowledge. This may limit the interpretability of the fraud detection results as it is hard to
know which actual transaction attributes are most relevant for fraud detection. This can be a critical issue in
Page 1049
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026
financial and regulatory settings where explainable artificial intelligence (XAI), explainability and model
interpretability are required for regulatory compliance, audit and stakeholder trust.
Third, smaller hyperparameter space and 3-fold cross-validation were used instead of more exhaustive
configurations because of the limits of the Google Colab environment. Although this method guarantees an
effective experimentation, future research can investigate bigger search space and higher-fold cross-validation
to optimize the model performance even further. Fourth, the decision threshold of 0.70 in this study was also
chosen by looking at the Precision Recall curve and might not be generalized best to all deployment situations.
The threshold is a design parameter that balances Precision with Recall on ensemble models; varying cost
structure of different fraud detection systems can be operating with a different threshold. In future research,
optimization of the threshold should be investigated using formal cost-sensitive analysis or calibration. Fifth, a
standard k-nearest neighbor interpolation method was adopted when implementing SMOTE in the current study.
More advanced versions, such as Borderline-SMOTE, SMOTE-ENN, and SMOTE-CGAN, as investigated by
Cheah et al. (2023) and Du et al. (2024), can potentially be more effective in either creating more realistic
synthetic samples or in cleaning noisy borderline cases. Sixth, the research did not use the tools of deep learning
that have demonstrated potential as sequence-sensitive fraud detectors, including autoencoders and graph neural
networks (Du et al., 2024; Zhu et al., 2024). It will be a logical step to extend the comparative framework to
these approaches, and cost sensitive learning strategies, in the future.
CONCLUSION
This study gave an imbalance-sensitive assessment and hyperparameter optimization of three supervised
machine learning models, namely Logistic Regression, Random Forest, and XGBoost, to detect credit card
frauds using the European Credit Card Fraud Dataset. This work has used CRISP-DM process model (Wirth and
Hipp, 2000) as a lifecycle model and all experiments performed on Google Colaboratory with Python 3.10. The
important methodological innovations encompass the implementation of SMOTE with training data only to
avoid data leakage, the implementation of stratified 3-fold cross-validation to hyperparameterize in the limits of
Google Colab, and the use of AUPRC as the main evaluation metric.
The initial baseline results indicated that all models had performance drawbacks without imbalance correction
and threshold optimization and had a range of 0.672 to 0.796 on the AUPRC values. After the SMOTE
augmentation, hyperparameter tuning, and probability threshold optimization of 0.70, XGBoost had the highest
overall performance with AUPRC = 0.817, ROC-AUC = 0.970, and a balanced Precision = Recall = F1 = 0.81.
The second best was the Random Forest (AUPRC = 0.805) which has the highest Precision (0.93) and thus best
suited with less false positives. Although Logistic Regression had the best Recall (0.86), it had a low Precision
(0.10), which reduces its usefulness as deployed.
The results support various critical conclusions: (1) ensemble techniques, especially XGBoost and random forest
are more appropriate to severely imbalanced fraud detection tasks when compared to linear classifiers; (2) a
combination of SMOTE and probability threshold tuning can address the Precision-Recall trade-off that cannot
be resolved by over-sampling alone; (3) AUPRC is a more informative indicator than ROC-AUC or model
accuracy Further studies should address formal cost-sensitive threshold optimization, enhanced variants of
SMOTE, higher-fold cross-validation, and use of deep learning techniques to enhance further the performance
of fraud detection.
REFERENCES
1. Bergstra, J., & Bengio, Y. (2012). Random Search for Hyper-Parameter Optimization. Journal of
Machine Learning Research, 13(10), 281–305.
https://www.jmlr.org/papers/volume13/bergstra12a/bergstra12a.pdf
2. Bisong, E. (2019). Google Colaboratory. In Building machine learning and deep learning models on
Google Cloud Platform: A comprehensive guide for beginners (pp. 59–64). Apress.
https://doi.org/10.1007/978-1-4842-4470-8_7
Page 1050
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026
3. Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32.
https://doi.org/10.1023/A:1010933404324
4. Carneiro, T., Da NóBrega, R. V. M., Nepomuceno, T., Bian, G.-B., De Albuquerque, V. H. C., &
Reboucas Filho, P. P. (2018). Performance analysis of Google Colaboratory as a tool for accelerating
deep learning applications. IEEE Access, 6, 61677–61685.
https://doi.org/10.1109/ACCESS.2018.2874767
5. Chawla, N. V., Bowyer, K. W., Hall, L. O., & Kegelmeyer, W. P. (2002). SMOTE: Synthetic minority
over-sampling technique. Journal of Artificial Intelligence Research, 16, 321–357.
https://doi.org/10.1613/jair.953
6. Cheah, P. C. Y., Yang, Y., & Lee, B. G. (2023). Enhancing financial fraud detection through addressing
class imbalance using hybrid SMOTE-GAN techniques. International Journal of Financial Studies, 11(3),
Article 110. https://doi.org/10.3390/ijfs11030110
7. Chen, T., & Guestrin, C. (2016). XGBoost: A scalable tree boosting system. Proceedings of the 22nd
ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 785–794.
https://doi.org/10.1145/2939672.2939785
8. Chung, J., & Lee, K. (2023). Credit card fraud detection: An improved strategy for high recall using
KNN, LDA, and linear regression. Sensors, 23(18), Article 7788. https://doi.org/10.3390/s23187788
9. Creswell, J. W., & Creswell, J. D. (2018). Research design: Qualitative, quantitative, and mixed methods
approaches (5th ed.). SAGE Publications.
10. Dantas, R. M., Firdaus, R., Jaleel, F., Mata, P. N., Mata, M. N., & Li, G. (2022). Systemic acquired
critique of credit card deception exposure through machine learning. Journal of Open Innovation:
Technology, Market, and Complexity, 8(4), Article 192. https://doi.org/10.3390/joitmc8040192
11. Davis, J., & Goadrich, M. (2006). The relationship between Precision-Recall and ROC curves.
Proceedings of the 23rd International Conference on Machine Learning, 233–240.
https://doi.org/10.1145/1143844.1143874
12. Du, H., Lv, L., Guo, A., & Wang, H. (2023). AutoEncoder and LightGBM for credit card fraud detection
problems. Symmetry, 15(4), Article 870. https://doi.org/10.3390/sym15040870
13. Du, H., Zhang, Y., Li, X., & Wang, Q. (2024). A novel method for detecting credit card fraud problems.
PLOS ONE, 19(3), Article e0294537. https://doi.org/10.1371/journal.pone.0294537
14. Fernández, A., García, S., Herrera, F., & Chawla, N. V. (2018). SMOTE for learning from imbalanced
data: Progress and challenges, marking the 15-year anniversary. Journal of Artificial Intelligence
Research, 61(1), 863–905. https://doi.org/10.1613/jair.1.11192
15. Herland, M., Bauder, R. A., & Khoshgoftaar, T. M. (2018). Approaches for identifying U.S. medicare
fraud in provider claims data. Health Care Management Science, 23(1), 2–19.
https://doi.org/10.1007/s10729-018-9460-8
16. Ileberi, E., Sun, Y., & Wang, Z. (2021). Performance evaluation of machine learning methods for credit
card fraud detection using SMOTE and AdaBoost. IEEE Access, 9, 165286–165294.
https://doi.org/10.1109/ACCESS.2021.3134330
17. Ileberi, E., Sun, Y., & Wang, Z. (2022). A machine learning based credit card fraud detection using the
GA algorithm for feature selection. Journal of Big Data, 9, Article 24. https://doi.org/10.1186/s40537-
022-00573-8
18. Nilson Report. (2022). Card fraud losses worldwide [Issue 1209]. Nilson Report. https://nilsonreport.com
19. Saito, T., & Rehmsmeier, M. (2015). The precision-recall plot is more informative than the ROC plot
when evaluating binary classifiers on imbalanced datasets. PLOS ONE, 10(3), Article e0118432.
https://doi.org/10.1371/journal.pone.0118432
20. Schroer, J., Kruse, F., & Gómez, J. M. (2021). A systematic literature review on applying CRISP-DM
process model. Procedia Computer Science, 181, 526–534. https://doi.org/10.1016/j.procs.2021.01.199
21. Snoek, J., Larochelle, H., & Adams, R. P. (2012). Practical Bayesian Optimization of Machine Learning
Algorithms. arXiv. https://doi.org/10.48550/arxiv.1206.2944
22. Strelcenia, E., & Prakoonwit, S. (2023). Improving classification performance in credit card fraud
detection by using new data augmentation. AI, 4(1), 172–198. https://doi.org/10.3390/ai4010008
23. Tripathy, N., Nayak, S. K., Godslove, J. F., Friday, I. K., & Dalai, S. S. (2022). Credit card fraud detection
using logistic regression and synthetic minority oversampling technique (SMOTE) approach.
Page 1051
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026
International Journal of Computer and Communication Technology, 8(4), 38–45.
https://doi.org/10.47893/ijcct.2022.1438
24. ULB Machine Learning Group. (2013). Credit card fraud detection [Dataset]. Kaggle.
https://www.kaggle.com/datasets/mlg-ulb/creditcardfraud
25. Wirth, R., & Hipp, J. (2000). CRISP-DM: Towards a standard process model for data mining.
Proceedings of the 4th International Conference on the Practical Applications of Knowledge Discovery
and Data Mining, 29–39. http://cs.unibo.it/~danilo.montesi/CBD/Beatriz/10.1.1.198.5133.pdf
26. Zhu, M., Zhang, Y., Gong, Y., Xu, C., & Xiang, Y. (2024). Enhancing credit card fraud detection: A
neural network and SMOTE integrated approach. Journal of Theory and Practice of Engineering Science,
4(02), 23–30. https://doi.org/10.53469/jtpes.2024.04(02).04
