Tractable Explanations for d-DNNF Classifiers

Compilation into propositional languages finds a growing number of practical uses, including in constraint programming, diagnosis and machine learning (ML), among others. One concrete example is the use as classifiers, one natural question how to explain predictions made. This paper shows that for classifiers represented with some best-known languages, different kinds explanations can be computed polynomial time. These include deterministic decomposable negation normal form (d-DNNF), so any language strictly less succinct than d-DNNF. Furthermore, describes optimizations, specific Sentential Decision Diagrams (SDDs), which are shown yield more efficient algorithms practice.