MAIRE - A Model-Agnostic Interpretable Rule Extraction Procedure for Explaining Classifiers

The paper introduces a novel framework for extracting model-agnostic human interpretable rules to explain classifier’s output. rule is defined as an axis-aligned hyper-cuboid containing the instance which classification decision has be explained. proposed procedure finds largest (high coverage) such that high percentage of instances in have same class label being explained precision). Novel approximations coverage and precision measures terms parameters are defined. They maximized using gradient-based optimizers. quality rigorously analyzed theoretically experimentally. Heuristics simplifying generated explanations achieving better interpretability greedy selection algorithm combines local creating global model covering large part space also proposed. agnostic, can applied any arbitrary classifier, all types attributes (including continuous, ordered, unordered discrete). wide-scale applicability validated on variety synthetic real-world datasets from different domains (tabular, text, image).