Abstract

The continued advances in artificial intelligence (AI), including those in machine learning (ML), raise concerns regarding their deployment in high-risk and safety-critical domains. Motivated by these concerns, there have been calls for the verification of systems of AI, including their explanation. Nevertheless, tools for the verification of systems of AI are complex, and so error-prone. This paper describes one initial effort towards the certification of logic-based explainability algorithms, focusing on monotonic classifiers. Concretely, the paper starts by using the proof assistant Coq to prove the correctness of recently proposed algorithms for explaining monotonic classifiers. Then, the paper proves that the algorithms devised for monotonic classifiers can be applied to the larger family of stable classifiers. Finally, confidence code, extracted from the proofs of correctness, is used for computing explanations that are guaranteed to be correct. The experimental results included in the paper show the scalability of the proposed approach for certifying explanations.