Formalizing fuzzy explainability: enhancing XAI trustworthiness with fuzzy-ontology-based explanatory properties

Professionals across various application domains of AI have long requested increased transparency of predictions generated by AI black-box models. In response, eXplainable AI (XAI) has gained popularity with researchers developing techniques to provide trustworthy explanations. We believe that incorporating expert knowledge into XAI systems through ontologies is a reliable approach. The Ontology-Based Image classifier (OBIC) introduced a flexible architecture compatible with any machine learning model, using OWL2 ontologies to represent expert knowledge. Later work extended OBIC by introducing ”explanatory” properties linked to logical features deduced from the datasets. This enables explanations using both explicit and implicit interpretable elements, independent of specific data types. This study formalizes ”explanatory” properties by integrating them with fuzzy logic. Building on previous research, we introduce new characteristics to address the vagueness of real-world data and shared knowledge. By embedding these properties into a fuzzy OWL2 ontology, we advance trustworthiness through fuzzy membership degrees, which serve as certainty levels for explanations, particularly in fuzzy knowledge environments.