Borrower Credit Risk Assessment based on Gated Transformer and Enhanced Mapping Gated Multi-Layer Perception

Borrower credit risk assessment plays a critical role in the stability of financial institutions, yet the accuracy of existing models is often constrained by the mixed processing of categorical and continuous features, which differ fundamentally in data characteristics. To address this issue, this paper proposes a novel credit risk assessment model named GTEMG, which integrates a Gated Transformer (Gated-Trans) and an Enhanced Mapping Gated Multi-Layer Perceptron (EM-gMLP). The methodology consists of four steps. First, categorical features are embedded into dense vectors, while continuous features are normalized. Second, a GatedTrans module with a Gaussian Gating Unit (GaGU) extracts high-order nonlinear interactions among categorical features. Third, the resulting context embeddings are concatenated with normalized continuous features. Fourth, an EM-gMLP module with Joint Feature Mapping (JFM) captures both sample-wise and feature-wise interaction information. Finally, a linear layer outputs the credit risk score. Experiments are conducted on three benchmark datasets—German Credit (GC), Australian Credit Approval (AC), and Japanese Credit (JC)—with evaluation metrics of Area Under the Curve (AUC) and Kolmogorov–Smirnov (KS) statistic. The proposed GTEMG model achieves AUC and KS values of 85.49% and 62.42% on GC, 98.31% and 94.59% on AC, and 90.28% and 69.82% on JC, respectively, outperforming state-of-the-art tabular data models. It is demonstrated the proposed method is effective and reliable in credit risk assessment.