Multi-view reject inference for semi-supervised credit scoring with consistency training and three-way decision

In credit scoring, reject inference based on semi-supervised learning has shown better performance compared to those based on statistical methods. However, the problem of inconsistent data distribution between accepted and rejected samples still exists during model training, which may violate the smoothness assumption of semi-supervised learning. Besides, multi-view learning has demonstrated its effectiveness, but its validity in reject inference still needs to be verified. Therefore, this paper proposes a multi-view reject inference approach (MRIA) based on three-way decision and consistency training. Specifically, with the aid of three-way decision, we sift valuable rejected samples from the profitability and accuracy objects, which brings the rejected samples better approximate the smooth assumption of semi-supervised learning. Then, based on the above-mentioned two objects, we construct multi-views by utilizing feature selection and train the reject inference model using consistency training, which can enhance the reliability and robustness. Finally, a dynamic fusion method built on the distance to model (DM) is employed for multi-view fusion. This paper not only theoretically demonstrates that high-quality data augmentation consistency training can result in a smaller error bound for the reject inference tasks, but also verifies the effectiveness of MRIA via a series of experimental analysis.