Conceptual
Browse Instruct Curate

Approximate Probabilistic Sufficient Reasons for Linear Models

A relaxation in formal explainable AI, the (delta, epsilon)-sufficient reason, that makes probabilistic feature-subset explanations tractable. Students learn how sufficient reasons certify a classifier's decision, why exact and delta-probabilistic versions are computationally hard or inapproximable even for simple models, and how adding an epsilon tolerance yields an explanation that can be computed efficiently for linear classifiers.