ExplAIner: A Declarative Query Language for Explaining Classification Models

Explanation methods need standardization: The paper identifies that the AI explainability field has developed many different explanation approaches without a unified framework, making it difficult to compare, combine, or systematically apply these methods in practice.

Current explanation tools have significant limitations: The researchers found that FOIL, an existing interpretability query language, cannot express important types of explanations (specifically "optimality-based" explanations that identify the best or most minimal explanations), highlighting gaps in current explanation tools available to practitioners.

ExplAIner provides a unified framework: The new query language can express multiple explanation types—including abductive explanations (why something happened), contrastive explanations (why this prediction instead of that one), feature-based explanations, and distance-based explanations—in a single, consistent system.

Computational complexity remains a challenge: While ExplAIner advances the field, the paper proves that computing explanations involves complex computational problems that are inherently difficult to solve, which has practical implications for how quickly explanation systems can operate on real models.

Optimization matters for practical explanations: The paper introduces Opt-FOIL to find explanations that are "minimal" (the simplest or most concise), recognizing that practitioners need not just any explanation, but the most useful and understandable ones.