On Tractable XAI Queries based on Compiled Representations

Abstract

One of the key purposes of eXplainable AI (XAI) is to develop techniques for understanding predictions made by Machine Learning (ML) models and for assessing how much reliable they are. Several encoding schemas have recently been pointed out, showing how ML classifiers of various types can be mapped to Boolean circuits exhibiting the same input-output behaviours. Thanks to such mappings, XAI queries about classifiers can be delegated to the corresponding circuits. In this paper, we define new explanation and/or verification queries about classifiers. We show how they can be addressed by combining queries and transformations about the associated Boolean circuits. Taking advantage of previous results from the knowledge compilation map, this allows us to identify a number of XAI queries that are tractable provided that the circuit has been first turned into a compiled representation.

On Efficiently Explaining Graph-Based Classifiers

“…Our work can be related with recent work on bayesian classifiers and decision graphs (Shih, Choi, and Darwiche 2018;Shih, Choi, and Darwiche 2019;Darwiche and Hirth 2020), but also languages from the knowledge compilation (KC) map (Audemard, Koriche, and Marquis 2020;Audemard et al 2021). In addition, we build on the recent results on the interpretability and the need for explainability of DTs (Izza, Ignatiev, and Marques-Silva 2020).…”

“…see (Marques-Silva et al 2020)). Similarly, other recent work (Audemard, Koriche, and Marquis 2020) investigates languages from the knowledge compilation map, which consider binary features. In addition, the tractable classifiers considered in (Audemard, Koriche, and Marquis 2020) for AXps do not intersect those studied in this paper.…”

“…In addition, the tractable classifiers considered in (Audemard, Koriche, and Marquis 2020) for AXps do not intersect those studied in this paper. In a companion work, (Audemard et al 2021) prove that for several XAI queries proposed in (Audemard, Koriche, and Marquis 2020), including AXp extraction, there exist polynomial algorithms for the case of DTs. In (Barceló et al 2020)…”

“…In contrast, formal approaches to XAI have been proposed in recent years (Shih, Choi, and Darwiche 2018;Ignatiev, Narodytska, and Marques-Silva 2019a;Shih, Choi, and Darwiche 2019;Ignatiev, Narodytska, and Marques-Silva 2019b; Darwiche and Hirth 2020;Audemard, Koriche, and Marquis 2020;Audemard et al 2021) (albeit it can be related to past work on logic-based explanations (e.g. (Shanahan 1989;Falappa, Kern-Isberner, and Simari 2002;Pérez and Uzcátegui 2003))).…”

“…Despite these drawbacks, it is plain that PI-explanations offer a sound basis upon which one can expect to develop theoretically sound and practically effective approaches for computing explanations. For example, more recent work has demonstrated the tractability of PI-explanations for some ML models (Izza, Ignatiev, and Marques-Silva 2020;Audemard, Koriche, and Marquis 2020;Marques-Silva et al 2020;Marques-Silva et al 2021;, in some cases allowing for polynomial delay enumeration (Marques- Silva et al 2020). Also, recent work (Ignatiev 2020;Ignatiev and Marques-Silva 2021; showed that, even for ML models for which computing a PI-explanation is NPhard, scalability may not be an obstacle.…”