Work

Estimand-Relative Validity and Context-Driven Instability of Shapley Explanations for Tabular Foundation Models

Year: 2026
Type: preprint
Abstract: Shapley explanations for tabular foundation models depend on a removal operator that determines how predictions are evaluated when features are absent. Although different operators induce different cooperative games, their effects on attribution validity and stability in in-context tabular models remain insufficiently understood. We compare five removal operators for TabPFN v2 and TabICL v2 using exact Shapley enumeration, controlled synthetic data with known or high-precision reference attributions, real-world tabular datasets, and a preregistered factorial design. Operator performance was strongly estimand-dependent. Mean imputation and missing-value encoding most closely approximated interventional attributions, whereas recontextualisation and conditional estimation through in-context learning yielded the lowest errors for the conditional estimand; these rankings also varied with feature dependence. Across nuisance configurations, attribution variability was dominated by the sampled in-context examples, while column order, label coding, and model seed contributed little. A planted-null-feature control further showed that both models frequently changed their predictions in response to features constructed to contain no label information. The analysis is limited to two binary-classification foundation models and exact enumeration in low-dimensional settings, so the observed rankings should not be extrapolated to larger feature spaces or other model families. These findings show that Shapley explanations for tabular foundation models are not interpretable without specifying both the removal operator and the intended estimand, and that context sensitivity should be reported as part of explanation reliability.
Language: English
Open Access status: green
OpenAlex ID: W7225892219
Cites: 33
Cited by:
Related to: 10