Reproduce “Adapting to Latent Subgroup Shifts via Concepts and Proxies” 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
Ibrahim Alabdulmohsin, Nicole Chiou, Alexander D'Amour, Arthur Gretton, Sanmi Koyejo, Matt J. Kusner, Stephen R. Pfohl, Olawale Salaudeen and 2 more
The paper develops adaptation for unlabeled targets under latent subgroup shift. It combines source concepts and a noisy subgroup proxy with target features to infer changed subgroup prevalences and reweight subgroup-specific predictions. Under stated graph and rank conditions, structured Wasserstein autoencoder adaptation achieved target AUROC 0.8131 versus 0.6730 for an unstructured model in simulation, not real-world validation.
Key insights
- The optimal target predictor for labels given features, $q(Y|X)$, is non-parametrically identifiable under latent subgroup shifts, even with unobserved confounders, provided specific auxiliary concept and proxy variables are available.
- Standard domain adaptation techniques (covariate shift, label shift) generally incur non-zero error under latent subgroup shifts, highlighting the necessity for more advanced methods.
- The specific causal graph structure, including the roles of concepts as information bottlenecks and independent proxies for latent subgroups, is essential for achieving the proposed non-parametric identification.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.