On Tackling Explanation Redundancy in Decision Trees

Decision trees (DTs) epitomize the ideal of interpretability machine learning (ML) models. The decision motivates explainability approaches by so-called intrinsic interpretability, and it is at core recent proposals for applying interpretable ML models in high-risk applications. belief DT justified fact that explanations predictions are generally expected to be succinct. Indeed, case DTs, correspond paths. Since ideally shallow, so paths contain far fewer features than total number features, DTs succinct, hence interpretable. This paper offers both theoretical experimental arguments demonstrating that, as long equates with succinctness explanations, then ought not deemed introduces logically rigorous path explanation redundancy, proves there exist functions which must exhibit arbitrarily large redundancy. also only a very restricted class can represented no In addition, includes results substantiating redundancy observed ubiquitously trees, including those obtained using different tree algorithms, but wide range publicly available trees. proposes polynomial-time algorithms eliminating practice require negligible time compute. Thus, these serve indirectly attain irreducible,