Distribution-Aware Diffusion Learning for Robust Credit Risk Prediction under Data Drift

This study addresses common challenges in credit risk prediction, including data distribution shift, class imbalance, and noise interference, by proposing a robust adaptive prediction framework that integrates distribution correction with a diffusion-based generation mechanism. The model consists of a feature embedding layer, a distribution alignment module, a diffusion generation network, a temporal attention encoder, and an adaptive risk decoder. The feature embedding layer maps multi-source financial data into a unified latent space to reduce scale differences and structural deviations among heterogeneous features. The distribution alignment module employs a maximum mean discrepancy constraint to dynamically match source and target domain distributions, improving model stability across time and industry contexts. Based on this foundation, the diffusion generation mechanism enhances the robustness of feature representations through a stepwise denoising reconstruction process, effectively handling abnormal samples and noise contamination. The temporal attention module captures the evolving dependencies of risk over time, emphasizing key features that influence risk variation, while the adaptive risk decoder outputs stable and interpretable risk predictions. Experimental results show that the proposed model outperforms multiple deep learning baselines in accuracy, recall, and F1-score, demonstrating its robustness and generalization ability in dynamic credit environments. This research provides a unified modeling approach that balances distributional consistency and generative robustness, offering new methodological support for risk modeling in complex financial data.