[3Xin4] Poster session 1

[3Xin4-71] Counterfactual Explanation for Incomplete Data

〇Kentaro Kanamori¹, Takuya Takagi¹, Ken Kobayashi², Yuichi Ike³ (1.Fujitsu Limited, 2.Tokyo Institute of Technology, 3.The University of Tokyo)

Keywords:explainable machine learning, counterfactual explanation, missing value, submodular optimization

This paper proposes a new framework of counterfactual explanation (CE) that works even in the presence of missing values. CE is a post-hoc explanation method that provides an action for altering the prediction result of a machine learning model into the desired one. We first empirically and theoretically show the risk that imputation of missing values in an instance affects an action with respect to its validity, as well as the features that the action suggests changing. To alleviate such a risk and obtain good actions, we formulate a task of constructing a compact set of imputation-action pairs so that it includes at least one valid action at high probability whatever the values of missing features are. We also show that the task is a submodular maximization problem, which can be efficiently solved by a greedy algorithm. Experimental results on public datasets demonstrated the efficacy of our method.