Dempster-Shafer Fusion of Semi-supervised Learning Methods for Predicting Defaults in Social Lending

摘要

In social lending, it is hard to know whether borrowers will repay well or not. Most researchers use supervised learning for default prediction, but labeling data by hand is time-consuming. Moreover, labeling results of semi-supervised learning methods are not the same each other. In this paper, we propose a fusion method of label propagation and transductive SVM based on Dempster-Shafer theory for precisely labeling unlabeled data to improve the performance. We remove few unlabeled data with lower reliabilities in labeling results and fusion of the two results based on Dempster-Shafer theory. We have conducted experiments with supervised learning method trained with labeled unlabeled data. As a result, the proposed method produced the best accuracies, 6.15% higher than the result trained with labeled data only, and 1.3% higher than the conventional methods.