Metrics for Evaluating Actionability in Explainable AI

To enable recourse, explanations provided to people should be actionable, that is, explain what a person should do to change the model’s decision. However, what actionability means in the context of explainable AI is unclear. In this paper, we explore existing tools that others developed to evaluate actionability in their respective domains. To our knowledge, no prior work in the XAI field has developed such a tool to evaluate the actionability of explanation. We conducted an experimental study to validate two existing actionability tools for discriminating the actionability of two types of explanations. Our results indicate that the two existing actionability tools reveal metrics relevant for conceptualising actionability for the XAI community.