InstanceSHAP: an instance-based estimation approach for Shapley values

Abstract The growth of artificial intelligence applications requires to find out which explanatory variables mostly contribute to the predictions. Model-agnostic methods, such as SHapley Additive exPlanations (SHAP) can solve this problem: they can determine the contribution of each variable to the predictions of any machine learning model. The SHAP approach requires a background dataset, which usually consists of random instances sampled from the train data. In this paper, we aim to understand the insofar unexplored effect of the background dataset on SHAP and, to this end, we propose a variant of SHAP, InstanceSHAP, that uses instance-based learning to produce a more effective background dataset for binary classification. We exemplify our proposed methods on an application that concerns peer-to-peer lending credit risk assessment. Our experimental results reveal that the proposed model can effectively improve the ordinary SHAP method, leading to Shapley values for the variables that are more concentrated on fewer variables, leading to simpler explanations.