Researchers Decompose Uncertainty in SHAP Explanations

A new preprint (v2 posted 24 Feb 2026) proposes a method to decompose uncertainty in SHAP values for tree ensembles, separating aleatoric, epistemic, and entanglement components via Dempster-Shafer evidence theory and Dirichlet-process hypothesis sampling. The authors validate the approach across three real-world use cases and show that high-SHAP features can be unstable, advising data and model improvements.