Reproduce “A Theory of Label Propagation for Subpopulation Shift” on OpenResearch

About this paper

Tianle Cai, Ruiqi Gao, Jason D. Lee, Qi Lei

The paper introduces a theoretical framework for label propagation in domain adaptation settings characterized by subpopulation shift, providing error bounds for consistency-based methods. This framework demonstrates how unlabeled data can be used to transfer knowledge to target domains and improve classifier performance.

Key insights

• Label propagation through consistency regularization, coupled with a subpopulation-level expansion property, can effectively transfer knowledge to target domains, even if the initial teacher classifier performs poorly on the target.
• The theoretical bounds on target error depend on the ground truth consistency error, the domain shift ratio, and the teacher's margin on the source, rather than directly on the teacher's target accuracy.
• The generalized label propagation framework unifies various learning paradigms, including semi-supervised learning, domain expansion, and multi-source domain adaptation, under a common theoretical explanation.

Summary by alphaXiv.