L2loRe: a method for explaining the reject option

Research on human-computer interaction emphasise the importance of reliability in hybrid decision-making systems. Trust hinges on the performance and trustworthiness of AI, achievable through accuracy metrics, confidence scores, eXplainable AI, and abstention mechanisms. This study presents an explainable abstaining classifier named Learning to Reject via Local Rule-based Explanations (L2loRe), a novel approach that leverages the distance between data points and counterfactuals to evaluate the confidence of predictions, thus facilitating the formulation of a rejection policy and generating clear explanations for the reasoning behind predictions or rejections.