Efficient Explanations With Relevant Sets

Abstract

Recent work proposed δ-relevant inputs (or sets) as a probabilistic explanation for the predictions made by a classifier on a given input. δ-relevant sets are significant because they serve to relate (model-agnostic) Anchors with (model-accurate) PI-explanations, among other explanation approaches. Unfortunately, the computation of smallest size δ-relevant sets is complete for NP PP , rendering their computation largely infeasible in practice. This paper investigates solutions for tackling the practical limitations of δ-relevant sets. First, the paper alternatively considers the computation of subset-minimal sets. Second, the paper studies concrete families of classifiers, including decision trees among others. For these cases, the paper shows that the computation of subset-minimal δ-relevant sets is in NP, and can be solved with a polynomial number of calls to an NP oracle. The experimental evaluation compares the proposed approach with heuristic explainers for the concrete case of the classifiers studied in the paper, and confirms the advantage of the proposed solution over the state of the art.Preprint. Under review.

No silver bullet: interpretable ML models must be explained

“…Logic-based (or formal) explanation approaches have been studied in a growing body of research in recent years (Shih et al, 2018 , 2019 ; Ignatiev et al, 2019a , b , c , 2020a , 2022 ; Narodytska et al, 2019 ; Wolf et al, 2019 ; Audemard et al, 2020 , 2021 , 2022a , b ; Boumazouza et al, 2020 , 2021 ; Darwiche, 2020 ; Darwiche and Hirth, 2020 , 2022 ; Izza et al, 2020 , 2021 , 2022a , b ; Marques-Silva et al, 2020 , 2021 ; Rago et al, 2020 , 2021 ; Shi et al, 2020 ; Amgoud, 2021 ; Arenas et al, 2021 ; Asher et al, 2021 ; Blanc et al, 2021 , 2022a , b ; Cooper and Marques-Silva, 2021 ; Darwiche and Marquis, 2021 ; Huang et al, 2021a , b , 2022 ; Ignatiev and Marques-Silva, 2021 ; Izza and Marques-Silva, 2021 , 2022 ; Liu and Lorini, 2021 , 2022a ; Malfa et al, 2021 ; Wäldchen et al, 2021 ; Amgoud and Ben-Naim, 2022 ; Ferreira et al, 2022 ; Gorji and Rubin, 2022 ; Huang and Marques-Silva, 2022 ; Marques-Silva and Ignatiev, 2022 ; Wäldchen, 2022 ; Yu et al, 2022 ), and are characterized by formally provable guarantees of rigor, given the underlying ML models. Given such guarantees of rigor, logic-based explainability should be contrasted with well-known model-agnostic approaches to XAI (Ribeiro et al, 2016 , 2018 ; Lundberg and Lee, 2017 ; Guidotti et al, 2019 ), which offer no guarantees of rigor.…”