Symbolic Explanations for ML Models: A Formal Approach via FCA and Forgetting

Ponente: Víctor Ramos González

"The deployment of Machine Learning (ML) systems in decision-making pipelines has intensified the demand for explainability. A central challenge in eXplainable AI (XAI) is the tension between model fidelity and human usability: explanations must be faithful to the model's behavior while remaining concise enough for stakeholder communication and auditing. In this paper, we address this problem by framing explanation as a signature management task through the lens of Formal Concept Analysis (FCA) and Variable Forgetting.We introduce an enriched implicational language, AL+⟂, which extends traditional Attribute Logic to handle both positive dependencies and mutual exclusion constraints—essential for modeling one-hot encodings and classification labels. We propose the use of variable forgetting (conservative retraction) as a formal explanation operator. This allows for the projection of the complex internal logic of a model onto a user-selected vocabulary, effectively 'abstracting away' irrelevant internal variables while preserving all logical entailments over the chosen attributes. We demonstrate the effectiveness of this framework by applying it to Binary Neural Networks (BNNs). By constructing neural contexts, we extract certified, human-readable implications that describe both global model behavior and local predictions. Our results show that this logic-based approach not only provides high-fidelity explanations but also uncovers structural redundancies within the network. This work provides a principled bridge between sub-symbolic representations and the rigorous requirements of formal Knowledge Representation."

Más información.