Reproduce “Explaining the Uncertain: Stochastic Shapley Values for Gaussian Process Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Siu Lun Chau, Krikamol Muandet, Dino Sejdinovic

The authors extend Shapley attribution for Gaussian processes (GPs) so explanations capture uncertainty in model predictions as well as feature credit. They derive means and covariances for stochastic Shapley values; BayesGP-SHAP also includes uncertainty from coalition estimation. In experiments, a Shapley-prior GP predicts held-out diabetes explanations with lower RMSE than two predictors, within the reported setup.

Key insights

  • By treating explanations as random variables, Stochastic Shapley Values (SSVs) provide a richer, more informative understanding of feature contributions by quantifying their uncertainty.
  • The uncertainty of SSVs correctly correlates with model uncertainty, increasing in regions with sparse training data.
  • The correlation structure of SSVs can reveal hidden relationships and dependencies between features in the model.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.