A Three-stage Data Mining Model for Reject Inference

Reject inference is a term that distinguishes attempts to correct models in view of the characteristics of rejected applicants. The main difficulty in establishing reject inference model is that the ´through-the-door´ applicant population is unavailable. In this paper, we propose a hybrid data mining technique for reject inference. It is a three-stage approach: k-means cluster, support vector machines classification and computation of feature importance. By combining the samples of the accepted applicants and the new applicants, we obtain representative samples. To some extent, this is cost-free. Analytic results demonstrate that our method improves the predictive performance while still retaining interpretability.