A Few Good Counterfactuals: Generating Interpretable, Plausible and Diverse Counterfactual Explanations

Abstract

Counterfactual explanations are an important solution to the Explainable AI (XAI) problem, but good, “native” counterfactuals can be hard to come by. Hence, the popular methods generate synthetic counterfactuals using “blind” perturbation, by manipulating feature values to elicit a class change. However, this strategy has other problems, notably a tendency to generate invalid data points that are out-of-distribution or that involve feature-values that do not naturally occur in a given domain. Instance-guided and case-based methods address these problems by grounding counterfactual generation in the dataset or case base, producing synthetic counterfactuals from naturally-occurring features, and guaranteeing the reuse of valid feature values. Several instance-guided methods have been proposed, but they too have their shortcomings. Some only approximate grounding in the dataset, or do not readily generalise to multi-class settings, or are limited in their ability to generate alternative counterfactuals. This paper extends recent case-based approaches by presenting a novel, general-purpose, case-based solution for counterfactual generation to address these shortcomings. We report a series of experiments to systematically explore parametric variations on common datasets, to establish the conditions for optimal performance, beyond the state-of-the-art in instance-guided methods for counterfactual XAI.