A method that extends formal explainable AI to LightGBM gradient-boosted classifiers by encoding the trained model as a Boolean formula and using a SAT solver to compute minimal, provably-correct feature sets responsible for each prediction. It introduces class-level explanations that characterize model behavior across an entire predicted class rather than a single instance, and a formal-explanation-driven procedure for both generating and detecting adversarial examples, contrasting the determinism of these guarantees with the instability of heuristic attribution methods such as SHAP and LIME.
XReason is a formal explainable-AI tool that encodes a trained tree-ensemble classifier as a Boolean formula and uses a SAT solver to compute minimal, provably-correct feature sets (abductive explana…