Relevance Metric for Counterfactuals Selection in Decision Trees

Explainable Machine Learning is an emerging field in the Machine Learning domain. It addresses the explicability of Machine Learning models and the inherent rationale behind model predictions. In the particular case of example-based explanation methods, they are focused on using particular instances, previously defined or created, to explain the behaviour of models or predictions. Counterfactual-based explanation is one of these methods. A counterfactual is an hypothetical instance similar to an example whose explanation is of interest but with different predicted class. This paper presents a relevance metric for counterfactual selection called sGower designed to induce sparsity in Decision Trees models. It works with categorical and continuous features, while considering number of feature changes and distance between the counterfactual and the example. The proposed metric is evaluated against previous relevance metrics on several sets of categorical and continuous data, obtaining on average better results than previous approaches.