“Better” Counterfactuals, Ones People Can Understand: Psychologically-Plausible Case-Based Counterfactuals Using Categorical Features for Explainable AI (XAI)

Abstract

A recent surge of research has focused on counterfactual explanations as a promising solution to the eXplainable AI (XAI) problem. Over 100 counterfactual XAI methods have been proposed, many emphasising the key role of features that are “important” or “causal” or “actionable” in making explanations comprehensible to human users. However, these proposals rest on intuition rather than psychological evidence. Indeed, recent psychological evidence [22] shows that it is abstract feature-types that impact people’s understanding of explanations; categorical features better support people’s learning of an AI model’s predictions than continuous features. This paper proposes a more psychologically-valid counterfactual method, one extending case-based techniques with additional functionality to transform feature-differences into categorical versions of themselves. This enhanced case-based counterfactual method, still generates good counterfactuals relative to baseline methods on coverage and distances metrics. This is the first counterfactual method specifically designed to meet identified psychological requirements of end-users, rather than merely reflecting the intuitions of algorithm designers.