Quantitatively Evaluating the Validity of Contrastive Generators for Recourse

Ian Howell, Eleanor Quint, Hongfeng Yu

Abstract: Explanations of classifiers can provide recourse to those impacted, i.e., the ability to facilitate those impacted to understand and potentially change the classification. Contrastive explanations provide this recourse by producing an alternative input close to the original such that the label is changed to the desired, implying actions in the difference between the original and contrastive inputs. However, quantitatively evaluating contrastive explanations remains a challenging task. In particular, some state-of-the-art contrastive explanation algorithms for decisions made by deep neural networks can produce inputs that are out-of-distribution or adversarial and are thus either infeasible for a user to achieve or do not change the label according to the underlying data distribution, respectively. Past work has termed contrastive examples valid if they cross the decision boundary of the classifier. However, this definition does not encompass these failure modes and thus is not suitable to evaluate contrastive methods when used for recourse. In this paper, we define a new type of validity, called distributional validity, that checks for these failure modes. We experiment with the distributional validity of state-of-the-art contrastive explanation methods and find that the best contrastive method depends on the architecture of the classification model.