Reproduce “$ϕ$-test: Global Feature Selection and Inference for Shapley Additive Explanations” on OpenResearch

About this paper

Dongseok Kim, Hyoungsun Choi, Mohamed Jismy Aashik Rasool, Gisung Oh

The ϕ-test framework integrates Shapley attributions with selective inference to generate statistically disciplined global feature-importance tables for black-box models. It quantifies feature impact with post-selection p-values and confidence intervals, achieving high predictive fidelity with sparser, more stable, and robust explanations across various model architectures and data resamples.

Key insights

• Initial SHAP-based feature screening can be treated as an ancillary step, allowing subsequent selective inference to validly condition on the selected candidate feature set.
• By formalizing the feature selection in a linear surrogate model as a polyhedral event, established selective inference theory can be applied to provide valid post-selection p-values and confidence intervals for black-box model explanations.
• The ϕ-test transforms qualitative SHAP-based importance rankings into quantitative statistical statements, enabling a more rigorous assessment of feature effects while accounting for the bias introduced by data-driven selection.

Summary by alphaXiv.