Temporal-Attentive Graph Neural Network withHierarchical Feature Fusion for Credit DefaultPrediction

Accurate credit default prediction is fundamental to effective risk management and sustainable lending in modern financial systems. While traditional statistical models provide interpretability, they struggle with nonlinear temporal dynamics and interdependencies characteristic of contemporary borrower behavior. This paper introduces the Temporal-Attentive Graph Neural Network with Hierarchical Feature Fusion (TAG-HFF), a unified deep learning framework that simultaneously models temporal evolution and relational structure in credit data through innovative architectural components. Our architecture integrates a temporal-attentive mechanism for emphasizing salient behavioral episodes, a graph neural network for encoding borrower relationships, and hierarchical feature fusion preserving semantic organization of financial attributes. To address class imbalance challenges prevalent in credit datasets, we employ SMOTE (Synthetic Minority Over-sampling Technique) during preprocessing. Comprehensive empirical evaluation using rigorous 5-fold cross-validation demonstrates superior performance across all evaluation metrics. TAG-HFF achieves an AUC of 0.844 with statistical significance confirmed by non-parametric testing (Wilcoxon signed-rank test $p < 0.05$ vs. 4/7 comparative models, Friedman test $\chi^2 = 26.06$, $p = 0.0002173$). Statistical analysis confirms TAG-HFF's dominance across discrimination, calibration, and robustness metrics. Beyond quantitative superiority, interpretability analyses using feature importance and SHAP values reveal strong alignment between model decisions and established financial principles. The proposed framework enables more informed credit decisions, substantial reduction in default-related losses, and optimized portfolio risk-return balance, positioning TAG-HFF as the superior choice for modern credit risk assessment and regulatory compliance.