Abstract
Fraudulent use of payment cards poses a serious
legal violation, severely damaging personal credit profiles and
banking entities. Detecting fraudulent credit card activity remains a persistent issue for financial systems due to the low
occurrence rate of fraud and the necessity for swift, reliable
identification. Conventional detection techniques tend to be inefficient, demanding significant manual effort and time. This study
presents a novel framework that integrates an advanced boosting
algorithm with data balancing strategies to enhance classification
effectiveness on a skewed dataset. An extensive evaluation was
conducted using labeled transaction records to assess the efficacy of various classification models, including Support Vector
Machine (SVM), Gaussian Naïve Bayes (GNB), Decision Trees
(DT), Adaptive Boosting Regression (ABR), Logistic Regression
(LR), and XGBoost. The research also explores how performance
metrics are influenced by oversampling, feature refinement, and
hyperparameter adjustments. The results show that the SVM,
ABR, LR, GNB, DT, and XGBoost classifiers have an overall
accuracy (OA) of 99.95%, 99.92%, 99.95%, 97.89%, 99.93%, and
99.96%, respectively. Findings indicate that the proposed hybrid
model, particularly when employing the XGBoost algorithm with
synthetic data expansion, delivers superior results compared to
standard classifiers. Metrics such as precision, recall, and F1-
score consistently favor this method, highlighting its advantage
in identifying rare fraudulent cases with remarkable accuracy.
