Evidential Counterfactual Explanations for Instance-Based Classification under the Belief Function Framework

Explainability is now an essential requirement for deploying
artificial intelligence systems in sensitive environments. While neighborhood based models such as k-NN are naturally interpretable, they remain limited in their ability to handle uncertainty. The Evidential k-NN (EKNN) classifier, based on the Dempster–Shafer theory, introduces belief func tions that explicitly model uncertainty between neighbors. However, this interpretative potential remains underexploited in the context of counterfactual explanations. To address this gap, we propose in this paper EKNN-NICE, a new counterfactual explanation method for EKNN that handles uncertainty through belief function theory, in order to ensure evidentially reliable explanations.