An ASP-Based Approach to Counterfactual Explanations for Classification

Abstract

We propose answer-set programs that specify and compute counterfactual interventions as a basis for causality-based explanations to the outcomes from classification models. They can be applied with black-box models, and also with models that can be specified as logic programs, such as rule-based classifiers. The main focus is on the specification and computation of maximum-responsibility counterfactual explanations, with responsibility becoming an explanation score for features of entities under classification. We also extend the programs to bring into the picture semantic or domain knowledge. We show how the approach could be extended by means of probabilistic methods, and how the underlying probability distributions could be modified through the use of constraints.

A Survey of Contrastive and Counterfactual Explanation Generation Methods for Explainable Artificial Intelligence

“…They assess how discriminatory the generated counterfactual explanations are for the given classification task output. On the other hand, Bertossi redefines the concept of causal explanation [67]. Following a causal account of contfactual explanation, Fernández et al introduce weakly causal irreducible counterfactual explanation [136].…”

“…Following Woodward's theory, Schneider and Rohlfing define a counterfactual as ''a theoretically relevant manipulation of the observed case in order to ascertain whether this manipulation would make a difference to the outcome'' [116]. Further, Bertossi defines a causal counterfactual explanation to be a set of the original feature values in the given data instance that are affected by a minimal counterfactual intervention [67] (where minimality is assumed to be based on a partial order relation on counterfactual interventions). Conversely, Andreas and Casini reconsider explanatory counterfactuals to be ''hypothetical assumptions about the values of quantities or the values of propositions''.…”

“…Thus, the outlined classification-oriented frameworks are evaluated on classifiers based on logistic regression [55], [136], [153], [158], decision trees [46], [80], [122], [140], [150], [155], [159], gradient boosted decision trees [147], support vector machines [131], [138], [146], random forests [81], [86], [142]- [144], neural networks [6], [48], [49], [91], [129], [130], [133], [135], [139], [141], [145], [148], [151], or combinations of these [100], [105], [134], [152], [154], [160]. In three studies [67], [128], [137], the classifiers used in the experiments are not specified.…”

“…Numerical values (or intervals of values) associated to the most relevant features usually explain the behavior of AI algorithms. They can be represented as logical formulas [67], [93], [94] or in tabular form reflecting necessary changes to affect the decision [55], [56], [61], FIGURE 11. Numbers of frameworks grouped by output representation with respect to the type of contfactual explanation generated.…”

“…Aguilar-Palacios et al [61] refer to Lipton's definition of contrastive explanation [12]. Referring to Pearl [37], Bertossi redefines causal explanation in the context of XAI to be ''a set of feature values for the entity under classification that is most responsible for the outcome'' [67].…”