Sufficient Reasons for Classifier Decisions in the Presence of Domain Constraints

Recent work has unveiled a theory for reasoning about the decisions made by binary classifiers: classifier describes Boolean function, and reasons behind an instance being classified as positive are prime-implicants of function that satisfied instance. One drawback these works is they do not explicitly treat scenarios where underlying data known to be constrained, e.g., certain combinations features may exist, observable, or required disregarded. We propose more general theory, also based on prime-implicants, tailored taking constraints into account. The main idea view classifiers describing partial functions undefined instances satisfy constraints. prove this simple results in parsimonious reasons. That is, account (e.g., ignoring, them negative instances) subsumed take illustrate improved succinctness synthetic learnt from real data.