Hybrid Contrastive Learning With Attention-Based Neural Networks for Robust Fraud Detection in Digital Payment Systems

Abstract

Fraud detection in digital payment systems is a critical challenge due to the growing complexity of transaction patterns and the inherent class imbalance in datasets. This article proposes a novel Hybrid Contrastive Learning framework integrating Siamese Networks with Attention-Based Neural Networks to effectively distinguish fraudulent from legitimate transactions. The proposed model achieves state-of-the-art results, surpassing 10 recent methods in key metrics, with a recall of 95.42%, precision of 97.35%, and ROC-AUC of 98.78% on the Credit Card Fraud Detection dataset. Cross-dataset evaluations using a simulated transaction dataset demonstrate consistent generalization, achieving a recall of 95.12% and ROC-AUC of 98.60%. An ablation study underscores the impact of attention mechanisms and contrastive learning, with the combined approach improving F1-score by up to 2.64%. Additionally, SHAP-based analysis reveals the importance of key features such as transaction amount and PCA-derived components in model decisions, enhancing explainability. The results establish the proposed framework as a robust, interpretable, and scalable solution for fraud prevention in digital payment systems.