Research map / Label-efficient and robust learning / Domain adaptation and generalization
Link-based Contrastive Learning for One-Shot Unsupervised Domain Adaptation
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.
Proposes link-based contrastive learning (LCL) for one-shot unsupervised domain adaptation, using in-domain must/cannot-links and bidirectionally matched cross-domain cluster links, evaluated on four UDA benchmarks and a surveillance face recognition dataset.
Problem
Unsupervised domain adaptation when only one labeled source example per category is available while the target domain is abundant and unlabeled (one-shot UDA), where existing UDA, few-shot UDA and source-free DA methods degrade sharply.
Approach
Builds pairwise must-link/cannot-link constraints from target clustering for discriminative feature learning, and aligns source and target clusters via bidirectional cluster matching that generates cross-domain links between samples and prototypes of the other domain.
Key results
- LCL reaches 47.25% average accuracy on Office-Home versus 30.45% for the Direct baseline and 40.74% for the best prior method PCS.
- On the VSA surveillance face recognition task, LCL improves average accuracy over Direct by 11.84% across six face recognition backbones (36.20% vs 21.62% average).
- LCL reaches 74.34% average on Office-31 and 63.09% on VisDA-2017 under the one-shot setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LCL instead of Direct | OSUDA adaptation · VisDA-2017 | accuracy: 40.20 -> 63.09 | Helped |
| Use LCL instead of Direct | OSUDA adaptation · Office-31 (Avg.) | accuracy: 61.08 -> 74.34 | Helped |
| Use LCL instead of Direct | OSUDA adaptation · Office-Home (Avg.) | accuracy: 30.45 -> 47.25 | Helped |
| Use LCL instead of Direct | OSUDA adaptation · DomainNet (Avg.) | accuracy: 15.34 -> 27.65 | Helped |
| Use LCL instead of Direct | face recognition backbones · VSA (Avg.) | accuracy: 21.62 -> 36.20 | Helped |
| Add prototypical contrastive learning (LPC) (ablation) | LCL on Office-Home · Office-Home (Avg.) | accuracy: 30.45 -> 37.13 | Helped |
| Add in-domain link-based contrastive learning (LILC) (ablation) | LCL on Office-Home · Office-Home (Avg.) | accuracy: 37.13 -> 42.45 | Helped |
| Add cross-domain link-based contrastive learning without bidirectional matching (ablation) | LCL on Office-Home · Office-Home (Avg.) | accuracy: 42.45 -> 44.82 | Helped |
| Use bidirectional matching for cross-domain links (ablation) | LCL on Office-Home · Office-Home (Avg.) | accuracy: 44.82 -> 47.25 | Helped |
Surprising findings
- Existing UDA, few-shot UDA and source-free DA methods often perform worse than the Direct no-adaptation baseline in the one-shot setting, e.g., UCDIR drops 20.58% on Office-31 and SDAT drops 9.34% on Office-Home.
Limitations
- Clustering errors can yield unreliable in-domain links.
- Bidirectional matching may be too strict, leaving many cross-domain pairs unlinked.
- Results on the VSA face dataset are low in absolute terms (best 44.7% on PartialFC), and the domain gap can be too large for the method.
Experimental setup
Datasets: VisDA-2017 Office-31 Office-Home DomainNet VSA
Benchmarks: VisDA-2017 Office-31 Office-Home DomainNet VSA
Metrics: adaptation accuracy
Baselines: PCS UCDIR SDAT CAF ICON ProtoOT MCUDA SHOT NRC AaD SFDA-DE BMD
Builds on: PCS AaD SHOT MoCo-style prototype contrast
Closest papers by meaning
- Connect, Not Collapse: Explaining Contrastive Learning for Unsupervised Domain Adaptation (ICML 2022 · similarity 0.89)
- CLDA: Contrastive Learning for Semi-Supervised Domain Adaptation (NeurIPS 2021 · similarity 0.89)
- Contrastive Adaptation Network for Unsupervised Domain Adaptation (CVPR 2019 · similarity 0.88)
- SimProF: A Simple Probabilistic Framework for Unsupervised Domain Adaptation (AAAI 2025 · similarity 0.87)
- CASUAL: Conditional Support Alignment for Domain Adaptation with Label Shift (AAAI 2025 · similarity 0.87)
- COSDA: Counterfactual-based Susceptibility Risk Framework for Open-Set Domain Adaptation (ICML 2025 · similarity 0.87)
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.