Declarative Approaches to Counterfactual Explanations for Classification

摘要

We propose answer-set programs that specify and compute counterfactual interventions on entitiesthat are input on a classification model. In relation to the outcome of the model, the resultingcounterfactual entities serve as a basis for the definition and computation of causality-based explanationscores for the feature values in the entity under classification, namely responsibilityscores. The approach and the programs can be applied with black-box models, and also withmodels that can be specified as logic programs, such as rule-based classifiers. The main focusof this study is on the specification and computation of best counterfactual entities, that is,those that lead to maximum responsibility scores. From them one can read off the explanationsas maximum responsibility feature values in the original entity. We also extend the programsto bring into the picture semantic or domain knowledge. We show how the approach could beextended by means of probabilistic methods, and how the underlying probability distributionscould be modified through the use of constraints. Several examples of programs written in thesyntax of the DLV ASP-solver, and run with it, are shown.