Automated Evaluation of GNN Explanations with Neuro Symbolic Reasoning

Abstract

We propose using reasoners (typically used in semantic web and ontologies), to automatically evaluate the explanations from GNN explainability techniques, provided we are able to convert them into axioms. Typically a reasoner is fed the axioms (universal truth) along with the instance level axioms, and the goal. Then using several rules of inference such as modus ponens, modus tollens, rules of simplification, addition etc, the reasoner can generate the proof for the given goal. Optionally this proof in terms of concepts and relations can also be converted as natural language explanations.