A Survey of Counterfactual Explanations: Definition, Evaluation, Algorithms, and Applications

Explainable machine learning aims to reveal the reasons why black-box models make decisions. Counterfactual explanation is an example-based post-hoc explanation method. The counterfactual explanations aims to find the minimum perturbation that changes the model output with respect to the original instance. This study's goal is to review the literature that has already been written about counterfactual explanations and topics that are relevant to it. We provide a formal definition of counterfactual explanations and counterfactual explainer, and a summary and formulaic description of the properties of the generated counterfactual instances. In addition, we investigate the application of counterfactual explanations in two areas: model robustness, and generating feature importance. The findings demonstrate that the qualities necessary for counterfactual instances cannot be simultaneously satisfied by present methodologies. Finally, we go over potential future research directions.