Reproduce “Bayesian Invariance Modeling of Multi-Environment Data” 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

Luhuan Wu, Mingzhang Yin, Yixin Wang, John P. Cunningham, David M. Blei

BIP frames invariant feature selection as probabilistic inference about predictors with stable outcome relationships across environments. VI-BIP approximates subset uncertainty using factorized inclusion probabilities instead of enumerating subsets. In a 450-feature simulation, its exact recovery approaches the oracle as environments increase. On held-out yeast deletions, it yields 371 validated predictions, about 5.5 times ICP-s; this is not causal proof.

Key insights

  • The BIP posterior naturally concentrates on the true invariant feature set `z*` due to a maximized likelihood ratio `g(y | x_z*) / p_e(y | x_z*)` when `z` is indeed `z*`.
  • BIP quantifies uncertainty about invariant features through a posterior distribution, which can reveal multiple plausible invariant sets when data are insufficient to identify a unique one.
  • Theoretical guarantees establish posterior consistency, ensuring the model identifies true invariant features asymptotically, with contraction rates influenced by prior information and environment heterogeneity.

Summary by alphaXiv.

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