The Attribution Impossibility: No Feature Ranking Is Faithful, Stable, and Complete Under Collinearity

In three linesImpossibility theorem: no feature ranking can be simultaneously faithful, stable, and complete under collinearity. Authors quantify the result for 4 model classes, propose DASH (Diversified Aggregation of SHAP) as resolution, and formally verify 305 Lean 4 theorems. Consequence: 68% of public datasets exhibit attribution instability.

Summary generated by Claude — human-verified