Post-hoc explainable methods: uncertainty in counterfactual explanations

Abstract

With the introduction of the EU’s GDPR in 2017, lawmakers have emphasized the importance of algorithmic decisions being explainable to people affected by those decisions. This has led to a major focus being put on Explainable AI. A field which tries to build interpretable models which can explain their decision-making process. Alternatively, one can build a separate AI model that attempts to explain the internal process of another model. However, the explanations given by these methods are not actionable due to how they are generated. Counterfactual Explanations are a solution to this problem . These models aim to change the original input point in such a manner as to cause the new point to have the desired label.These new synthetic points are then presented to the user as alternate possibilities that they can try to to incorporate into their own feature vector and make the required changes .These explanations are more actionable for the user because they can see what changes led to a specific outcome. A drawback with this method of presenting counterfactuals to the user is that there is no good indicator for how many should be presented. The model might end up suggesting too few counterfactuals or too many leading to different problems respectively. Our research problem aims to rectify this by exploring whether it is possible to provide a range or distribution of values for the feature changes required and then present those to the user. We believe that this would be more beneficial to the user as it is easier to stay in a range than target a specific value. We will be highlighting this drawback of the way counterfactuals are presented to the user and expand upon our proposed research problem that we will be working on to fix this drawback.