Mitigating the Rashomon Effect in Counterfactual Explanation: A Game-theoretic Approach

Abstract

Counterfactual examples (CEs) are generally created to interpret the decision of a model. In this case, if a model makes a certain decision for an instance, the CEs of that instance reverse the decision of the model. There are many advantages of using counterfactuals as a way of explaining model decisions; however, there is one issue known as the Rashomon Effect that might dissuade target/intended users from using counterfactuals. If someone is presented with too many options, this might be overwhelming to them, and they might end up choosing an option that is not optimal or ideal for them. In this case, the Rashomon Effect is an impediment to realizing the full potential of counterfactual explanations. To utilize the full power of CEs and make them more helpful, we need to address the Rashomon Effect. In this work, we focus on this issue from game-theoretic perspectives to help target users make informed and feasible decisions by finding highly suitable CEs. In this case, finding good counterfactuals will be a game between two players where each of them tries to find better CEs.