A Temporal Causal Graph Approach for Credit Card Fraud Detection

Abstract

Credit card fraud has emerged as a significant and persistent issue in the financial sector, affecting consumers, businesses, and financial institutions globally. Traditional methods struggle with capturing complex temporal dependencies and causal relationships inherent in transactional data. To address these challenges, this study proposes a Temporal Causal Graph Contrastive Learning (TCGCL) framework to enhance the accuracy of credit card fraud detection. The TCGCL integrates temporal causal graphs with contrastive learning, aiming to improve detection performance significantly. The framework was evaluated against the state-of-the-art Pick and Choose Graph Neural Network (PC-GNN) method using metrics such as accuracy, precision, recall, and F1-score. Experiments were conducted on datasets of varying sizes (10,000, 50,000, and 100,000 instances) and across different epochs (10, 50, and 100 epochs). Findings reveal that the TCGCL achieved an average accuracy improvement of 3% over PC-GNN, reaching 95.8% with the largest dataset and highest number of epochs. The precision of TCGCL improved by approximately 2.5%, recording a peak value of 96.3%, TCGCL demonstrated a recall enhancement of around 3%, with a maximum of 95.5%.The framework attained the highest F1-score of 95.9%, outperforming PC-GNN by 2.8%.The results indicate that TCGCL effectively models temporal causal relationships and leverages contrastive learning to distinguish between fraudulent and legitimate transactions. The framework's performance improved significantly with larger datasets and extended training epochs, highlighting its robustness and scalability for real-world applications. Future work will focus on real-time implementation, hyperparameter optimization, and enhancing model explains ability to further solidify TCGCL's potential as a leading solution in fraud detection.