The Limits of Feature Ranking: A Major Breakthrough in Explainable AI

Published on May 22, 2026

Traditional methods in feature ranking have relied on the belief that they can provide stable, faithful, and complete outputs. However, new research reveals a critical flaw: in the presence of collinear features, no ranking can maintain these three qualities simultaneously. This discovery challenges long-held assumptions in the field.

The study, published on arXiv, demonstrates that when features become collinear, ranking reduces to arbitrary outcomes, akin to a coin flip. Researchers quantified this impossibility across four model classes and proposed a solution—DASH, or Diversified Aggregation of SHAP. This ensemble method promises stability, even in cases where features are symmetrically relevant.

Analyzing 77 public datasets, the researchers found that 68% displayed attribution instability, raising serious concerns about existing SHAP-based methods. The findings suggest that switching to alternative approaches, like conditional SHAP, fails to address the core issue of collinearity. This necessitates a reevaluation of how feature importance is assessed in machine learning contexts.

The implications of this research extend to fairness auditing, revealing that SHAP-based discrimination audits may be fundamentally flawed under collinearity conditions. The work culminates in a set of practical diagnostic tools and a formally verified impossibility theorem, marking a significant milestone in explainable AI. This pivotal breakthrough redefines our understanding of feature ranking and its limitations.

Related News