Interpretable Feature Selection Using Local Information for Credit Assessment

[ Information ]

nIPS 2018 Workshop on Challenges and Opportunities for AI in Financial Services: the Impact of Fairness, Explainability, Accuracy, and Privacy, 2018.

[ Authors ]

Sangwoong Yoon, Yonho Song, Minsoo Kim, Frank C. Park, and Yung-Kyun Noh

[ Abstract ]

Data analysis is interpretable when a small subset of features are used. However, the selection of useful subset is difficult because the pattern of associations between the target and the input variables are different across clusters. In this paper, we try to identify such associations correctly using neighborhood information for estimating the locally usefulness of features. We show that a plug-in of density estimators cancels the leading terms of finite-sampling bias of Jensen-Shannon divergence estimator, and a straightforward application of this estimator works well for capturing locally important features. In our experiments with FICO HELOC dataset, the proposed feature selection method reveals that different features are useful for explaining the risk of default loan in different groups of loan applicants.