Actionable Feature Discovery in Counterfactuals using Feature Relevance Explainers

Counterfactual explanations focus on “actionable knowledge” to help end-users understand how a Machine Learning model outcome could be changed to a more desirable outcome. For this purpose a counterfactual explainer needs to be able to reason with similarity knowledge in order to discover input dependencies that relate to outcome changes. Identifying the minimum subset of feature changes to action a change in the decision is an interesting challenge for counterfactual explainers. In this paper we show how feature relevance based explainers (i.e. LIME, SHAP), can inform a counterfactual explainer to identify the minimum subset of “actionable features”. We demonstrate our DisCERN (Discovering Counterfactual Explanations using Relevance Features from Neighbourhoods) algorithm on three datasets and compare against the widely used counterfactual approach DiCE. Our preliminary results show that DisCERN to be a viable strategy that should be adopted to minimise the actionable changes