Reject Inference Using Discriminative Dual Stack Sparse Auto-Encoders for Consumer Credit Risk Evaluation

Credit risk evaluation has gained substantial attention within financial institutions, serving as a pivotal tool to predict borrower repayment behavior and provide precise credit risk estimations. Traditional credit risk approaches discarded rejected applicants and were built only on accepted applicants, which posed sample selection bias issue. Previous reject inference methods solved the bias issue by incorporating information of rejected applicants. However, these methods assumed that the accepted and rejected samples had identical dimensions. In practical financial scenarios, financial institutions often encounter situations where the dimensions of accepted samples were larger than those of the rejected samples. Therefore, the additional features in accepted samples might not be fully utilized in the previous reject inference. In this study, we proposed a discriminative dual stack sparse auto-encoder (DD-SSAE) reject inference method that was suitable for the real scenarios. The proposed DD-SSAE has the following characteristics: (1) rejected samples were filtered based on our selection mechanism; (2) a stack sparse auto-encoder (SSAE), within a self-taught learning framework, was carried out to incorporate information of the selected rejected samples into the common features of accepted samples; and (3) a data fusion module, consisting of another SSAE network and a data fusion layer, was introduced to combine extra features with common features for accepted samples. The proposed method was verified on a Chinese consumer dataset and the findings illustrated its superiority over four conventional credit scoring models and five previous reject inference models.