Research on Credit Risk Identification Method for Supply Chain Finance Based on Heterogeneous Graph and Dynamic Distillation

Accurately identifying default risk among small and medium-sized enterprises (SMEs) in supply chain finance is challenging due to complex inter-firm relationships, severe class imbalance, and the need for interpretability. We propose a lightweight credit risk assessment framework that combines structure-aware embeddings from a heterogeneous supply-chain graph using Graph Contrastive Learning (GCL) with a distilled student classifier based on Dynamic-Temperature Knowledge Distillation (DKD-MLP). Class rebalancing (SMOTETomek) and feature pruning (RFE) further enhance training stability. On real-world SME and ChiNext board data from the Shenzhen Stock Exchange (2018–2023), the model achieves an AUC of 0.912, PR-AUC of 0.558, F1-score of 0.625, and Recall of 0.592, outperforming baselines such as GCN and GraphMLP. Ablation confirms the value of GCL and DKD, while SHAP analysis highlights the default history of core firms and accounts-receivable turnover as dominant predictors. The framework improves minority-class identification while maintaining efficiency and transparency, making it suitable for practical financial decision-making.