On the (Complete) Reasons Behind Decisions

Abstract Recent work has shown that the input-output behavior of some common machine learning classifiers can be captured in symbolic form, allowing one to reason about these using techniques. This includes explaining decisions, measuring robustness, and proving formal properties by reasoning corresponding classifiers. In this work, we present a theory for unveiling reasons behind decisions made Boolean study its theoretical practical implications. At core our is notion complete reason, which viewed as necessary sufficient condition why decision was made. We show how used computing notions such (also known PI-explanations abductive explanations), it determining classifier bias evaluating counterfactual statements “a will stick even if ...because ... .” linear-time algorithm decision, assuming represented circuit appropriate form. then computed answer many queries linear or polynomial time. finally conclude with case illustrates various techniques introduced.