PREF-XAI: Preference-Based Personalized Rule Explanations of Black-Box Machine Learning Models

Abstract

Explainable artificial intelligence (XAI) has largely focused on the objective dimension of explanations, producing model-centric outputs intended for a general user. Yet explanations should be relevant and acceptable to a particular user, whose goals, interests, background knowledge, and decision context shape their perceived usefulness. Although XAI research has considered personalization, existing approaches rely on heuristic adaptations or implicit user modeling and lack a principled framework for representing and learning individual preferences. We introduce Preference-Based Explainable Artificial Intelligence (PREF-XAI), which reframes the selection of explanations as a preference-driven decision problem. Our methodology combines rule-based explanations with formal preference learning. A user ranks a small subset of candidate explanatory rules, and this sparse feedback is used to infer an additive utility model that ranks the complete set of induced rules. The utility function weights are not treated as observable psychological quantities that users can explicitly provide. Instead, they are model parameters inferred from holistic comparisons of a few candidate rules. PREF-XAI therefore avoids requiring users to assign numerical weights to rule attributes, reducing cognitive effort and facilitating preference elicitation. Experiments on real-world datasets show that PREF-XAI can accurately reconstruct user preferences from limited feedback and identify relevant rules, including rules outside the initial comparison set. The approach connects XAI with preference learning and opens new avenues for research on interactive, personalized explanation systems.