Reproduce “Domain Adaptation with Conditional Distribution Matching and Generalized Label Shift” 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

Remi Tachet, Han Zhao, Yu-Xiang Wang, Geoff Gordon

This research introduces Generalized Label Shift (GLS) to address the performance degradation of adversarial domain adaptation (ADA) methods when source and target label distributions differ. It proposes importance-weighted adversarial learning algorithms that reweight source samples and achieve consistent accuracy improvements, outperforming baseline ADA methods by up to 9.3% on tasks with significant label distribution mismatches.

Key insights

  • Standard adversarial domain adaptation methods are fundamentally flawed under label shift, as aligning overall feature distributions can lead to worse joint source-target error by making detrimental trade-offs.
  • Achieving conditional invariance in a learned feature space ($P(Z|Y)$) is more practical and effective than in the raw input space for robust domain adaptation.
  • Minimizing the balanced error rate on the source domain while aligning a reweighted source feature distribution with the target feature distribution is a principled approach to overcoming label shift.

Summary by alphaXiv.

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