ASTERYX : A model-Agnostic SaT-basEd appRoach for sYmbolic and score-based eXplanations

ASTERYX provides two complementary types of explanations: The approach generates both symbolic explanations (sufficient reasons and counterfactuals) and numerical scores, offering users multiple ways to understand why a model made a particular decision.

Model-agnostic design increases practical applicability: By encoding any machine learning model into a symbolic representation, ASTERYX can work across different types of models rather than being limited to specific architectures, making it potentially useful across diverse AI applications in practice.

Addresses the black-box problem through symbolic reasoning: The method converts complex machine learning models into equivalent symbolic representations that can be analyzed logically, helping demystify how models arrive at their predictions—a key concern for responsible AI deployment.

Combines interpretability with relevance scoring: Beyond just explaining decisions, the approach assigns scores to explanations and features based on their relevance, helping users prioritize which factors matter most and distinguish between important and peripheral reasons for a prediction.

Experimental validation demonstrates feasibility: The authors show through experiments that their approach successfully generates both types of explanations, suggesting it could be practically implemented to improve transparency in real-world AI systems.