Sparse-enhanced additive interaction neural network for interpretable credit decision

Intelligent credit decision systems are crucial for financial institutions’ risk management, aiming to mitigate credit risk. While deep learning models offer high predictive accuracy, their opacity hinders decision support. Neural Additive Models (NAMs) offer feature-level interpretability but fail to capture complex interactions among credit risk factors. To enhance both accuracy and interpretability, we propose the Sparse-Enhanced Additive Interaction Neural Network (SAINTNet) for explainable credit scoring. SAINTNet advances NAM’s framework with dual-node additive modules and adaptive sparse feature selection, enabling autonomous feature learning. Leveraging entmax sparsity and optimized temperature settings, SAINTNet: (1) maintains interpretability, particularly for credit feature interactions; (2) achieves superior accuracy compared to black-box models. Experiments on four credit datasets demonstrate SAINTNet’s superior performance and systematic interpretability through global feature importance, local decision analysis, and interaction visualization, improving decision audits in high-risk credit scenarios.