Benchmark Evaluation of Counterfactual Algorithms for XAI: From a White Box to a Black Box

Counterfactual explanations have recently been brought to light as a potentially crucial response to obtaining human-understandable explanations from predictive models in Explainable Artificial Intelligence (XAI). Despite the fact that various counterfactual algorithms have been proposed, the state of the art research still lacks standardised protocols to evaluate the quality of counterfactual explanations. In this work, we conducted a benchmark evaluation across different model agnostic counterfactual algorithms in the literature (DiCE, WatcherCF, prototype, unjustifiedCF), and we investigated the counterfactual generation process on different types of machine learning models ranging from a white box (decision tree) to a grey-box (random forest) and a black box (neural network). We evaluated the different counterfactual algorithms using several metrics including proximity, interpretability and functionality for five datasets. The main findings of this work are the following: (1) without guaranteeing plausibility in the counterfactual generation process, one cannot have meaningful evaluation results. This means that all explainable counterfactual algorithms that do not take into consideration plausibility in their internal mechanisms cannot be evaluated with the current state of the art evaluation metrics; (2) the counterfactual generated are not impacted by the different types of machine learning models; (3) DiCE was the only tested algorithm that was able to generate actionable and plausible counterfactuals, because it provides mechanisms to constraint features; (4) WatcherCF and UnjustifiedCF are limited to continuous variables and can not deal with categorical data.