DTD-VAE: Disentangled Temporal Dependencies VAE for Credit Risk Prediction
This study proposes the DTD-VAE model to address the challenge of distinguishing credit-risk-related temporal patterns from general customer behavior patterns. By disentangling temporal dependencies and improving feature generation, the model enhances credit risk prediction accuracy, outperforming existing methods on multiple real-world datasets.
01 ABSTRACT
The paper introduces DTD-VAE, an enhanced VAE model for credit risk prediction. It employs an autoregressive temporal dependency learning mechanism and an element-wise gating mechanism to disentangle risk-specific features from general preferences. Experiments on six real-world datasets show significant performance improvements over baselines in ROC-AUC and Accuracy Ratio. The authors believe the model effectively captures temporal information, but the specific data sources and generalization should be considered.
02 KEY FINDINGS
- DTD-VAE disentangles temporal dependencies to distinguish risk-related features from general behavior patterns.
- Autoregressive temporal dependency learning enhances understanding of data structure.
- Element-wise gating enables fine-grained disentanglement, improving risk prediction.
- On six real-world datasets, ROC-AUC improves by 3.2%-4.86% and AR by 6.41%-9.71%.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV Q-FIN