Research map / Label-efficient and robust learning / Domain adaptation and generalization
Universal Semi-Supervised Domain Adaptation by Mitigating Common-Class Bias
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.
Introduces Universal Semi-Supervised Domain Adaptation (UniSSDA) and a prior-guided pseudo-label refinement (PGPR) strategy that uses a supervised head's per-instance class-group priors to counter common-class bias.
Problem
Semi-supervised domain adaptation assumes matching label spaces, while universal DA assumes unlabeled target data; in the combined UniSSDA setting, models and naive pseudo-labeling are biased toward common classes and misclassify target-private classes.
Approach
A supervised classifier head trained only on labeled data supplies per-instance priors over class groups (common, source-private, target-private); the semi-supervised head's predictions are reweighted to match these group priors and then averaged with the supervised predictions to produce refined pseudo-labels.
Key results
- On Office-Home with ResNet-34, the proposed method reaches 70.7% overall, versus 67.6% for the best baseline AdaMatch.
- On DomainNet-126 with ResNet-34, the method reaches 68.6% overall, outperforming the second-best method by 5.7%.
- With frozen CLIP ViT-L/14 on VisDA, overall accuracy is 88.0% versus 85.9% for UniOT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PGPR instead of the strongest SSDA baseline | UniSSDA with ResNet-34 · Office-Home open-partial | accuracy: 64.2 ±0.7 -> 67.4 ±1.8 | Helped |
| Use PGPR instead of the strongest SSDA baseline | UniSSDA with ResNet-34 · DomainNet-126 open-set | accuracy: 54.1 ±0.1 -> 61.2 ±0.8 | Helped |
| Add group reweighted pseudo-label refinement (ablation) | PGPR on DomainNet-126 C → P · open-set | accuracy: 50.4 -> 56.8 | Helped |
| Add classifier aggregated refinement after group reweighting (ablation) | PGPR on DomainNet-126 C → P · open-set | accuracy: 56.8 -> 57.4 | Helped |
| Add classifier aggregated refinement after group reweighting (ablation) | PGPR on DomainNet-126 C → P · partial-set | accuracy: 78.7 -> 78.5 | Hurt |
| Add PGPR to CDAC (ablation) | CDAC on DomainNet-126 C → P · open-set | accuracy: 50.0 -> 49.9 | No clear effect |
| Add PGPR to CDAC (ablation) | CDAC on DomainNet-126 C → P · open-partial | accuracy: 47.2 -> 55.0 | Helped |
| Add PGPR to PAC (ablation) | PAC on DomainNet-126 C → P · open-set | accuracy: 46.8 -> 49.2 | Helped |
| Add PGPR to PAC (ablation) | PAC on DomainNet-126 C → P · partial-set | accuracy: 75.2 -> 77.7 | Helped |
| Add PGPR to AdaMatch (ablation) | AdaMatch on DomainNet-126 C → P · open-set | accuracy: 50.2 -> 53.6 | Helped |
| Add PGPR to AdaMatch (ablation) | AdaMatch on DomainNet-126 C → P · partial-set | accuracy: 73.9 -> 77.8 | Helped |
Surprising findings
- Existing SSDA methods CDAC and PAC often fail to improve over the supervised S+T baseline in partial-set and open-set settings.
- On frozen foundation models, UniDA methods perform similarly to or worse than the supervised baseline S+T in most cases.
Limitations
- The method relies on a supervised head trained on few labeled target samples (3-shot), whose priors may be weak with fewer labels.
- Evaluation is limited to image classification benchmarks.
Experimental setup
Datasets: Office-Home DomainNet VisDA
Benchmarks: Office-Home DomainNet-126 DomainNet-345 VisDA
Metrics: target accuracy averaged over 12 domain pairs common vs target-private class accuracy
Baselines: S+T CDAC PAC AdaMatch DANCE UniOT DRw ProML OVANet
Builds on: AdaMatch FixMatch OVANet CDAC PAC
Closest papers by meaning
- Gradient-Based Sample Selection for Black-Box Universal Domain Adaptation (AAAI 2025 · similarity 0.90)
- Enhancing Semi-supervised Domain Adaptation via Effective Target Labeling (AAAI 2024 · similarity 0.90)
- Tackling Dimensional Collapse toward Comprehensive Universal Domain Adaptation (ICML 2025 · similarity 0.90)
- Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.89)
- A Theory of Learning Unified Model via Knowledge Integration from Label Space Varying Domains (CVPR 2025 · similarity 0.89)
- A Unified Framework for Heterogeneous Semi-supervised Learning (CVPR 2025 · similarity 0.89)
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.