A Principled Approach for Defining and Comparing Variable Importance Measures

Many popular explainability methods lack rigorous foundations: The paper identifies that numerous variable importance measures (VIMs) used to explain AI models are "heuristic in nature" without formal statistical guarantees, raising concerns about their reliability for real-world decision-making.

Shapley values and similar methods can produce false positives: The research demonstrates that popular explainability techniques like Shapley values can incorrectly assign importance to features that provide "no additional information," potentially leading to spurious correlations being mistaken for genuine predictive signals.

A principled framework can prevent misleading explanations: The authors propose an axiomatic approach that ensures features truly contributing to predictions are distinguished from those that don't, helping practitioners avoid being misled by explanations that appear authoritative but lack statistical validity.

Explainability methods need standardized evaluation criteria: The paper advocates for a systematic pipeline—defining theoretical objectives, estimating them properly, and deriving statistical guarantees—to enable meaningful comparison between different VIMs, addressing the current fragmentation where methods lack clear, comparable standards.

Statistical rigor matters for trustworthy AI: By connecting variable importance to established statistical methods like conditional independence testing, the framework demonstrates that responsible AI explanation requires grounding explainability techniques in formal statistical theory rather than relying on intuitive but unvalidated heuristics.