Reject inference in credit scoring using a three-way decision and safe semi-supervised support vector machine

Reject inference is a credit scoring technique that can resolve sample selection bias, with several statistical and machine learning methods having been recently employed to infer the status of rejected samples. This paper proposed a new reject inference method based on a three-way decision and a safe semi-supervised support vector machine (S4VM) model. In credit evaluations, the accepted sample is labeled and the rejected sample is unlabeled. This paper used S4VM to model both the accepted and rejected samples for reject infer-ence. Because of the basic semi-supervised learning assumption that the accepted and rejected sample distributions are similar, this paper used a three-way decision method to filter the rejected samples to ensure the accepted and rejected sample distributions were closer. It was found that this method filtered out some rejected samples that were signif-icantly different from the accepted sample distribution, which reduced the interference in the S4VM low-density separator. The proposed method was verified in four experiments on Chinese credit loan data, with the results verifying the effectiveness of the proposed reject inference S4VM method. (c) 2022 Elsevier Inc. All rights reserved.

Automated summary

This paper proposes a new reject inference method based on a three-way decision and a safe semi-supervised support vector machine (S4VM) model to resolve sample selection bias in credit scoring. The proposed method is effective in reducing interference and improving reject inference performance.