CTH-RiskGNN

Abstract

Credit card fraud detection has become increasingly challenging because modern financial transactions are highly heterogeneous, temporally evolving, structurally connected, and severely imbalanced. To address these challenges, this paper proposes CTH-RiskGNN, a causal temporal heterogeneous graph neural network designed to improve credit card fraud detection by jointly modeling transaction relations, temporal risk evolution, causal stability, and fraud risk diffusion. The ablation analysis further confirms that heterogeneous relation modeling, temporal encoding, causal stabilization, risk-diffusion memory, gated fusion, and imbalance-aware optimization all contribute to the final performance improvement. These findings indicate that integrating causal temporal graph learning with heterogeneous risk diffusion can effectively overcome the limitations of independent transaction classification and provide a more robust framework for detecting concealed, evolving, and relational financial fraud.