Reproduce “Contrastive Learning for Unsupervised Domain Adaptation of Time Series” 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

Yilmazcan Ozyurt, Stefan Feuerriegel, Ce Zhang

A framework developed at ETH Zürich and LMU Munich, CLUDA, addresses domain shift in time series data by learning domain-invariant, label-preserving contextual representations through a combination of momentum contrastive learning and adversarial training. It achieves state-of-the-art accuracy on human activity recognition benchmarks, with up to an 18.9% increase on HAR, and improves mortality prediction AUROC on medical datasets, such as a 2.2% gain on MIMIC-IV over prior methods.

Key insights

  • Explicitly learning and aligning contextual, label-preserving representations across domains is more effective for time series UDA than methods that only align general feature distributions.
  • Combining momentum contrastive learning for robust intra-domain representation learning with a custom nearest-neighbor contrastive loss for inter-domain alignment significantly enhances cross-domain generalization.
  • The integrated framework, incorporating adversarial training, successfully minimizes domain discrepancy while retaining crucial class-specific information, leading to improved performance on unlabeled target domains.

Summary by alphaXiv.

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