Provably Stable Feature Rankings with SHAP and LIME (2401.15800v2)

Abstract: Feature attributions are ubiquitous tools for understanding the predictions of machine learning models. However, the calculation of popular methods for scoring input variables such as SHAP and LIME suffers from high instability due to random sampling. Leveraging ideas from multiple hypothesis testing, we devise attribution methods that ensure the most important features are ranked correctly with high probability. Given SHAP estimates from KernelSHAP or Shapley Sampling, we demonstrate how to retrospectively verify the number of stable rankings. Further, we introduce efficient sampling algorithms for SHAP and LIME that guarantee the $K$ highest-ranked features have the proper ordering. Finally, we show how to adapt these local feature attribution methods for the global importance setting.