Answer-Set Programs for Reasoning about Counterfactual Interventions and Responsibility Scores for Classification
This provides a declarative method for generating attribution-based explanations in classification, which is incremental as it applies existing answer-set programming to a specific AI interpretability task.
The paper tackled the problem of explaining classification outcomes by using answer-set programs to specify counterfactual interventions and compute responsibility scores, demonstrating the approach with a naive-Bayes classifier example.
We describe how answer-set programs can be used to declaratively specify counterfactual interventions on entities under classification, and reason about them. In particular, they can be used to define and compute responsibility scores as attribution-based explanations for outcomes from classification models. The approach allows for the inclusion of domain knowledge and supports query answering. A detailed example with a naive-Bayes classifier is presented.