Explaining Cautious Random Forests via Counterfactuals

Abstract

Cautious random forests are designed to make indeterminate decisions when tree outputs are conflicting. Since indeterminacy has a cost, it seems desirable to highlight why a precise decision could not be made for an instance, or which minimal modifications can be made to the instance so that the decision becomes a single class. In this paper, we apply an efficient extractor to generate determinate counterfactual examples of different classes, which are used to explain indeterminacy. We evaluate the efficiency of our strategy on different datasets and we illustrate it on two simple case studies involving both tabular and image data.