| Counterfactual explanations are a widely adopted approach for interpreting the decisions of machine learning models, mainly in the context of classification problems. A wide variety of techniques have been proposed for the classification task. In this work, we address the generation of counterfactuals for explaining clustering decisions. We focus on k-means and Gaussian clustering and tackle counterfactual generation by defining equivalent classification problems. More specifically, a linear classifier is defined in the k-means case and a quadratic discriminant classifier is defined in the Gaussian clustering case. In this way, widely used methods developed in the classification context can be employed for clustering. We also propose a way to increase the plausibility of the generated countefactuals by moving the cluster boundary towards the target cluster. Experimental results on synthetic and real datasets demonstrate the feasibility and effectiveness of our approach. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.