Local Explanations via Necessity and Sufficiency: Unifying Theory and Practice

Necessity and sufficiency provide a unified framework for XAI: The paper demonstrates that two fundamental logical concepts—necessity (a factor must be present for an outcome) and sufficiency (a factor alone can cause an outcome)—can unify many existing but seemingly different explanation methods in AI, providing the field with much-needed theoretical foundations.

Current XAI methods lack consistency: The authors highlight that explainable AI has grown rapidly without rigorous theoretical grounding, leading to inconsistent application of necessity and sufficiency concepts across different explanation techniques—a gap this work aims to address.

A new algorithm offers sound and complete explanations: The paper provides a formal algorithm for computing explanatory factors that is both "sound" (produces correct explanations) and "complete" (finds all valid explanations), offering a more reliable approach than ad-hoc methods currently in use.

The framework is practical and competitive: Despite its theoretical rigor, the approach demonstrates competitive or superior performance compared to state-of-the-art XAI methods on various tasks, suggesting that stronger theoretical foundations don't require sacrificing real-world effectiveness.

Context matters for explanations: The framework emphasizes that explanations must be computed "with respect to a given context," recognizing that the same AI decision may require different explanations depending on the specific circumstances—an important consideration for responsible AI deployment.