Publication

Abstract : Credit card fraud detection remains one of the most serious problems in financial security, as fraudulent patterns evolve rapidly, and most transactions are valid. This paper proposes a Hybrid Adaptive Graph Neural Network, a novel framework combining graph-based transaction modeling with adaptive feature learning for improving detection performance. Different from those machine learning-based approaches that assess each transaction independently, HAGNN models a transactional graph in which nodes represent both customers and transactions, while edges represent the temporal relationship, behavioral links, and transactional dependencies. The model leverages the adaptive attention to dynamically weigh the connectivity based on anomaly likelihoods, enabling more emphasis on suspicious activities. Besides, the self-supervised contrastive learning module further refines the representations through distinguishing normal and fraudulent transaction embeddings. Further performance optimization is achieved by dynamically adjusting detection thresholds with the aid of a reinforcement learning-driven decision module and striking an optimal trade-off between drastically reducing false alarms and maximizing fraud recall. On real-world credit card datasets, HAGNN achieves 12-15% improvements of precision and recall compared to state-of-the-art techniques and thereby significantly reduces potential financial losses. It gives evidence that integrating graph-based deep learning, adaptive attention, and self-supervised learning together contributes much for the current advanced credit card fraud detection.