Answer-Set Programs for Reasoning About Counterfactual Interventions and Responsibility Scores for Classification

Attribution-Scores and Causal Counterfactuals as Explanations in Artificial Intelligence

“…The direct application of the responsibility score, as in [48,6], works fine for explanation scores when features are binary, say taking values 0 or 1, [12,11]. However, when features have more than two values, it makes sense to extend the definition of the responsibility score.…”

Reasoning about Counterfactuals and Explanations: Problems, Results and Directions

“…For our examples with decision-trees [8,9] and with naive-Bayes classifiers [10], we have used the DLV system (and its extensions) [20] that implements the ASP semantics. The classifiers can be specified directly inside the CIP, or can be invoked as external predicates [8].…”

“…CIPs are very flexible in that one can easily add domain knowledge or domain semantics, in such a way that certain counterfactuals are not considered, or others are privileged. With CIPs, many kinds of changes on the specification that are of potential interest can be easily and seamlessly tried out on-the-fly, for exploration purposes [10,9]. All these changes and alternatives are much more difficult to implement with a purely procedural approach.…”

“…Actually, there have been some efforts to design and investigate query languages for explanations [29]. ASP offers a query language for this task, and as a part of the same system that does the reasoning and computation [10]. The investigation of its full potential (or shortcomings) for this tasks remains to be carried out.…”

“…The first generation of answer-sets programming, the one that is mostly used, is not probabilistic, and does not provide much support for probabilistic reasoning. With some difficulty, one can do probabilistic reasoning through numerical aggregations (as with naive-Bayes classifiers in [10]).…”