Traditional credit evaluation models often struggle to model long behavioral sequences and complex relational risks. To overcome these limitations, this study proposed a hybrid deep learning framework named Sparse Attention Transformer and Graph Neural Network (SAT-GNN). The framework incorporated a Sparse Attention Transformer (SAT) to capture dynamic risk patterns in ultra-long individual behavior sequences. It also integrated a Graph Attention Network (GAT) to model risk contagion within heterogeneous entity graphs. In addition, an adaptive feature fusion layer was designed to balance individual-level features and group-level topological representations through trainable weights. Experiments were conducted on the publicly available IEEE-CIS dataset. The proposed model achieved an AUC of 0.952 and an AUPRC of 0.835. Compared with the industry-standard model LightGBM, recall improved by 20.7%. Ablation studies confirmed the contribution of each core component. The average inference latency per sample was only 8.4 milliseconds. Overall, the proposed framework delivered high predictive accuracy and low latency in real-time financial decision-making scenarios. The study provides a robust and efficient solution for credit risk management in free trade ports.
Wu et al. (2026) studied this question.