Data-Balanced Machine Learning Approach for Secure Online Payment Fraud Recognition

Abstract

Online payment fraud causes significant financial losses and reduces user trust. Detecting fraud is challenging due to extreme class imbalance where fraudulent transactions are rare compared to genuine ones. This paper presents a data-balanced ML approach applying SMOTE oversampling to improve fraudulent transaction representation. Random Forest, Logistic Regression, and XGBoost models are trained on the balanced dataset. Evaluation demonstrates that XGBoost on balanced data achieves 97.2% accuracy with 0.96 F1-score and 0.99 AUC, significantly improving over unbalanced training (83.1% accuracy, 0.52 F1-score for fraud class). The approach demonstrates that data balancing is critical for building effective online payment fraud detection systems, with the balanced XGBoost model reducing false negatives by 78% compared to imbalanced training.