Research map / Label-efficient and robust learning / Semi-supervised learning
A Unified Framework for Heterogeneous Semi-supervised Learning
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 Heterogeneous Semi-Supervised Learning (HSSL), where labeled and unlabeled data come from different domains sharing classes, and proposes Uni-HSSL, a 2C-class model trained with WMA pseudo-labeling, cross-domain prototype alignment, and progressive inter-domain mixup.
Problem
Semi-supervised learning assumes labeled and unlabeled data share the same distribution; in practice they often come from heterogeneous domains with different label and class-feature distributions, and test data come from both domains.
Approach
Treat semantic classes in each domain as separate classes (a doubled 2C-class label space) so one model distinguishes both domain and class, combined with a weighted-moving-average pseudo-label update, cross-domain class-prototype contrastive alignment, and a progressively scheduled inter-domain mixup.
Key results
- Uni-HSSL reaches 87.5% average accuracy on Office-31 vs 84.3% for the best SSL baseline (SimMatch) and 83.0% for MCC+Sup.
- Uni-HSSL reaches 74.7% average on Office-Home and 93.1% average on VisDA, versus 70.5% (CDAN+Sup) and 92.1% (MCC+Sup) respectively.
- On ISIC-2019, Uni-HSSL achieves 75.4% average accuracy vs 72.2% for SimMatch, the best baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Uni-HSSL instead of the strongest reported baseline, MCC+Sup | HSSL on Office-31 · Office-31 | accuracy: 83.0 -> 87.5 | Helped |
| Use Uni-HSSL instead of the strongest reported baseline, MCC+Sup | HSSL on Office-Home · Office-Home | accuracy: 72.3 -> 74.7 | Helped |
| Use Uni-HSSL instead of the strongest reported baseline, CDAN+Sup | HSSL on VisDA · VisDA | accuracy: 92.1 -> 93.1 | Helped |
| Use Uni-HSSL instead of the strongest reported baseline, FixMatch | HSSL on ISIC-2019 · ISIC-2019 | accuracy: 71.3 -> 75.4 | Helped |
| Add WMA to pseudo-label updates (ablation) | Uni-HSSL · Office-31 | accuracy: 86.8 -> 87.5 | Helped |
| Add the labeled-set cross-entropy classification loss (ablation) | Uni-HSSL · Office-31 | accuracy: 81.7 -> 87.5 | Helped |
| Add the unlabeled-set pseudo-label classification loss (ablation) | Uni-HSSL · Office-31 | accuracy: 86.1 -> 87.5 | Helped |
| Add cross-domain prototype alignment (ablation) | Uni-HSSL · Office-31 | accuracy: 86.5 -> 87.5 | Helped |
| Add the inter-domain mixup component (ablation) | Uni-HSSL · Office-31 | accuracy: 86.1 -> 87.5 | Helped |
| Use progressive rather than simple inter-domain mixup (ablation) | Uni-HSSL · Office-31 | accuracy: 85.6 -> 87.5 | Helped |
Surprising findings
- Standard UDA methods augmented with a supervised classifier and domain classifier (MCC+Sup) drop to 58.7% on ISIC-2019, attributed to class imbalance sensitivity.
Limitations
- Assumes the labeled and unlabeled domains share the same C semantic classes.
- Hyperparameters of baselines taken from original papers; evaluation limited to image classification benchmarks.
Experimental setup
Datasets: Office-31 Office-Home VisDA-2017 ISIC-2019
Benchmarks: Office-31 Office-Home VisDA-2017 ISIC-2019
Metrics: mean classification accuracy standard deviation over 3 runs
Baselines: Supervised FlexMatch FixMatch SimMatch CDAN+Sup MCC+Sup BiAdapt
Builds on: FixMatch MixUp MCC BiAdapt CDAN
Closest papers by meaning
- Universal Semi-Supervised Domain Adaptation by Mitigating Common-Class Bias (CVPR 2024 · similarity 0.89)
- Erasing the Bias: Fine-Tuning Foundation Models for Semi-Supervised Learning (ICML 2024 · similarity 0.88)
- Enhancing Semi-supervised Domain Adaptation via Effective Target Labeling (AAAI 2024 · similarity 0.87)
- CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 2025 · similarity 0.87)
- Unlocking the Potential of Unlabeled Data in Semi-Supervised Domain Generalization (CVPR 2025 · similarity 0.87)
- Scaling Up Semi-supervised Learning with Unconstrained Unlabelled Data (AAAI 2024 · 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.