Adaptive Confidence-Aware Transformer–Graph Fusion Network for Robust Credit Fraud Detection

Currently, finding fraudulent transactions made with credit card systems is a major concern for the research community. The growth of online payment systems has increased transaction volumes. Attackers use increasingly sophisticated techniques, including identity theft, card cloning, and the compromise of online transactions, making fraud detection challenging for financial institutions. Traditional detection models rely on fixed rules and thresholds, which limit their ability to adapt to emerging fraud patterns; these systems are not flexible to deal with new and dynamic types of fraud. The literature shows growing interest in the development of datadriven intelligent systems for the analysis of transactional behaviors and the automatic recognition of suspicious activities. In this context, an adaptive confidence-aware transformer–graph fusion network model is proposed for reliable fraud detection in imbalanced financial transaction datasets. This study empirically investigates the two publicly available financial transaction datasets: ULB credit card and PaySim synthetic dataset which is a commonly used standard in the literature, examining fraud detection; these datasets have realistic distributional characteristics and anonymized variables. The graph-based fusion model is evaluated on PaySim, where it obtains 99% in accuracy, 96 % in F1-score, and 99 % in receiver operating characteristic-area under the curve (ROC-AUC). It improves recall and F1-score compared to individual transformer and Graph Sample and Aggregate (GraphSAGE) models and conventional machine-learning methods.