Research map / Label-efficient and robust learning / Domain adaptation and generalization
Theoretical Performance Guarantees for Partial Domain Adaptation via Partial Optimal Transport
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Derives generalization bounds for partial domain adaptation based on partial Wasserstein distance, yielding explicit source-loss weights, and proposes WARMPOT, a mini-batch PW-based PDA algorithm competitive with SOTA on Office-Home.
Problem
Partial domain adaptation (target label space is a subset of source label space) typically relies on heuristic source-sample weights and domain alignment terms lacking theoretical justification; negative transfer from outlier source classes degrades performance.
Approach
Generalization bounds on the target loss involving a partial Wasserstein distance between empirical source/target feature (or feature-label) distributions, where source weights arise constructively as marginals of the optimal partial transport plan; the resulting weighted, PW-regularized objective is minimized as WARMPOT.
Key results
- WARMPOT is competitive with recent PDA methods on Office-Home, achieving higher average accuracy than PWAN and similar to MPOT.
- Replacing ARPM's weights with WARMPOT weights (ARPM+our-weights) improves on the prior SOTA on Office-Home.
- On P->A, WARMPOT assigns only 6.04% of weight mass to outlier source samples although 58.41% of source samples are outliers.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace ARPM weights with WARMPOT weights | ARPM · Office-Home | test accuracy | Helped |
| Replace ARPM weights with WARMPOT weights | ARPM · ImageNet → Caltech | performance | Helped |
| Use WARMPOT weighting instead of MPOT weighting | PDA weighting schemes · Office-Home | performance | Helped |
| Use WARMPOT weighting instead of ARPM weighting | PDA weighting schemes · Office-Home | performance | Helped |
| Use WARMPOT weighting instead of BA³US weighting | PDA weighting schemes · Office-Home | performance | No clear effect |
| Use WARMPOT weighting instead of MPOT weighting | PDA weighting schemes · ImageNet → Caltech | performance | Helped |
| Use WARMPOT weighting instead of ARPM weighting | PDA weighting schemes · ImageNet → Caltech | performance | Helped |
| Use WARMPOT weighting instead of BA³US weighting | PDA weighting schemes · ImageNet → Caltech | performance | No clear effect |
| Vary β across (0,1] (ablation) | WARMPOT · ImageNet → Caltech | test accuracy: variations not exceeding 2% | No clear effect |
| Vary αmax within (0,0.1] (ablation) | WARMPOT · ImageNet → Caltech | test accuracy | No clear effect |
| Set αmax above 0.1 (ablation) | WARMPOT · ImageNet → Caltech | test accuracy | Hurt |
Surprising findings
- Weights from the PW-based bound match BA3US-level performance while being computed per mini-batch rather than every 500 iterations.
Limitations
- Exact minimization of the bounds is computationally prohibitive, so WARMPOT relies on approximations.
- Hyperparameters (alpha, beta, eta1, eta2) selected by parameter search; sensitivity to alpha above 0.1 degrades performance on ImageNet to Caltech.
- Bounds do not capture additional uncertainty/robustness losses used by ARPM.
Experimental setup
Datasets: Office-Home ImageNet Caltech-256
Benchmarks: Office-Home ImageNet to Caltech
Metrics: test accuracy
Baselines: MPOT BA3US ARPM PWAN
Builds on: MPOT PWAN ARPM BA3US Wasserstein distance domain adaptation bounds
Closest papers by meaning
- Prototypical Partial Optimal Transport for Universal Domain Adaptation (AAAI 2023 · similarity 0.92)
- Mapping conditional distributions for domain adaptation under generalized target shift (ICLR 2022 · similarity 0.91)
- Reliable Weighted Optimal Transport for Unsupervised Domain Adaptation (CVPR 2020 · similarity 0.89)
- Incorporating Importance Weighting in Optimal Transport Based Domain Alignment (ICML 2026 · similarity 0.89)
- Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.88)
- Distinguish Then Exploit: Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment (CVPR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.