Counterfactual Explanations May Not Be the Best Algorithmic Recourse Approach

DOI

Algorithmic recourse is a rapidly developing subfield in explainable AI (XAI) concerned with providing individuals subject to adverse high-stakes algorithmic outcomes with explanations indicating how to reverse said outcomes. While XAI research in the machine learning community doesn't confine itself to counterfactual explanations, its algorithmic recourse subfield does, adopting the assumption that the optimal way to provide recourse is through counterfactual explanations. Though there has been extensive human-AI interaction research on explanations, translating these findings to the algorithmic recourse setting is non-obvious due to meaningful problem setting differences, leaving the question of whether counterfactuals are the most optimal explanation paradigm for recourse unanswered. While intuitively satisfying, the prescriptive nature of counterfactuals makes them vulnerable to poor outcomes when circumstances unknown to the decision-making and explanation generating algorithms affect re-application strategies. With these concerns in mind, we designed a series of experiments comparing different explanation methods in the recourse setting, explicitly incorporating scenarios where circumstances unknown to the decision-making and explanation algorithms affect re-application strategies. In Experiment 1, we compared counterfactuals with reason codes, a simple feature-based explanation, finding that they both yield comparable re-application success, and that reason codes led to better user outcomes when unknown circumstances had a high impact on re-application strategies. In Experiment 2, we sought to improve on reason code outcomes, comparing them to feature attributions, a more informative feature-based explanation, but found no improvements. Finally, in Experiment 3, we aimed to improve on reason code outcomes with a multiple counterfactual explanation condition, finding that multiple counterfactuals led to higher re-application success but still resulted in comparatively worse user outcomes in the face of high impact unknown circumstances. Taken together, these findings call into question whether the standard counterfactual paradigm is the best approach for the algorithmic recourse problem setting.