Shapley values assign each feature an importance score based on its contribution to a model's prediction. The method is mathematically principled — it satisfies uniqueness, efficiency, symmetry, and null-player axioms. But it is also causally naive: it evaluates each feature by marginalizing over the others using their observational distribution, which encodes correlations regardless of whether those correlations are causal.
Note (2026-11-14; read against the paper's full text): the essay says cc-Shapley recovers the true causal contributions and that the rankings under the two methods disagree on synthetic and real data. The paper's claims are narrower: it shows theoretically that cc-Shapley removes the spurious association induced by collider bias, it requires knowledge of the causal structure of the data (it is not a method for discovering it), and its comparison on synthetic and real-world datasets reports nullification or reversal of associations relative to univariate feature importance when moving from observational to cc-Shapley. It also says that a feature with negative relevance is not automatically a suppressor, because a feature can have negative relevance by really causing the target to decrease, so the right reading of an attribution chart can vary from situation to situation. The essay's statement that features high under Shapley drop to zero and vice versa is stated more generally than anything I found; the paper shows such cases in specific causal graphs (for example its breakfast and diabetes example), and I could not tell from the text how often it happens in the real datasets.
Martin and Haufe (arXiv:2602.20396) show this produces collider bias and suppression in feature importance rankings. A feature that is statistically associated with the outcome only because it shares a common effect with another feature — a collider — gets assigned nonzero importance despite having no causal influence. A genuine cause whose effect is partially suppressed by a correlated non-cause gets its importance reduced.
cc-Shapley incorporates causal structural knowledge — which variables cause which — into the marginalization step. Instead of evaluating a feature's contribution against the observational distribution of other features, it evaluates against the interventional distribution. This eliminates spurious associations from collider structures and recovers the true causal contributions.
The importance rankings change substantially. Features that ranked high under standard Shapley drop to zero under cc-Shapley, and vice versa. On both synthetic and real datasets, the two methods disagree about which features matter. The disagreement is not noise — it is the difference between association and causation showing up in the importance scores.
The general observation: a mathematically principled measure can be causally wrong. Shapley values satisfy axioms about fair allocation of credit. They do not satisfy axioms about causal attribution. When the goal is to understand why a prediction was made — in order to intervene, debug, or regulate — association-based importance is misleading. The mathematical correctness of the method does not guarantee the correctness of the interpretation.