On Computing Probabilistic Abductive Explanations

The most widely studied explainable AI (XAI) approaches are unsound. This is the case with well-known model-agnostic explanation approaches, and it also based on saliency maps. One solution to consider intrinsic interpretability, which does not exhibit drawback of unsoundness. Unfortunately, interpretability can display unwieldy redundancy. Formal explainability represents alternative these non-rigorous one example being PI-explanations. PI-explanations important drawbacks, visible arguably their size. Recently, has been observed that (absolute) rigor be traded off for a smaller size, by computing so-called relevant sets. Given some positive δ, set S features δ-relevant if, when in fixed, probability getting target class exceeds δ. However, even very simple classifiers, complexity sets prohibitive, decision problem NPPP-complete circuit-based classifiers. In contrast earlier negative results, this paper investigates practical number used classifiers include Decision Trees (DTs), Naive Bayes Classifiers (NBCs), several families obtained from propositional languages. Moreover, shows that, practice, easy compute. Furthermore, experiments confirm succinct considered.