Research map / Label-efficient and robust learning / Domain adaptation and generalization
CASUAL: Conditional Support Alignment for Domain Adaptation with Label Shift
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 CASUAL, a conditional adversarial support alignment method for unsupervised domain adaptation under label shift that minimizes a conditional symmetric support divergence, with a new target risk bound and a training scheme.
Problem
Unsupervised domain adaptation where the label marginal differs between source and target (label shift), under which marginal feature distribution alignment degrades performance.
Approach
Align the supports of class-conditional feature distributions (conditional symmetric support divergence) via an adversarial discriminator on joint feature-prediction outer products using pseudo-labels, instead of marginal support or distribution alignment, with a new CSSD-based target risk bound motivating the objective.
Key results
- On USPS to MNIST, CASUAL reaches 95.6% average per-class accuracy versus 93.4% for the second-best method (OSTAR), and 88.3% at alpha=0.5.
- On STL to CIFAR, CASUAL averages 75.1%, above the next best DANN* at 73.2%.
- On VisDA-2017, CASUAL averages 69.8%, above MIC at 69.3%, and is best at severe shift alpha=0.5 with 63.2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add conditional entropy loss to conditional support alignment (ablation) | CASUAL · USPS→MNIST | per-class accuracy: 92.4 -> 95.1 | Helped |
| Add virtual adversarial loss to conditional support alignment and conditional entropy (ablation) | CASUAL · USPS→MNIST | per-class accuracy: 95.1 -> 95.6 | Helped |
| Use CASUAL instead of the strongest reported baseline | USPS→MNIST, α=0.5 · USPS→MNIST, α=0.5 | per-class accuracy: +3.6% | Helped |
| Use CASUAL instead of the strongest reported baseline | STL→CIFAR, α=0.5 · STL→CIFAR, α=0.5 | per-class accuracy: +1.8% | Helped |
| Use CASUAL instead of the strongest reported baseline | VisDA-2017, α=0.5 · VisDA-2017, α=0.5 | per-class accuracy: +0.7% | Helped |
Surprising findings
- Support alignment methods like ASA and sDANN underperform distribution alignment methods when there is no label shift, which the authors attribute to support alignment being an extreme relaxation of distribution alignment.
- Distribution-alignment baselines such as SDAT and MIC perform worse than No DA under severe label shift (alpha=0.5 on USPS to MNIST).
Limitations
- Smaller CSSD does not always translate to better target accuracy, since the bound contains unobservable terms.
- Relies on pseudo-labels, with error accumulation under large domain shift mitigated by entropy conditioning.
- Evaluated only on visual benchmarks; DomainNet results deferred to appendix.
Experimental setup
Datasets: USPS MNIST STL CIFAR-10 VisDA-2017 DomainNet
Benchmarks: USPS to MNIST STL to CIFAR VisDA-2017
Metrics: per-class accuracy average accuracy across label shift levels
Baselines: No DA DANN* CDAN* VADA SDAT MIC DALN IWDAN IWCDAN sDANN ASA PCT
Builds on: adversarial support alignment (ASA) integral measure discrepancy (IMD) virtual adversarial training DANN CDAN
Closest papers by meaning
- Can We Evaluate Domain Adaptation Models Without Target-Domain Labels? (ICLR 2024 · similarity 0.91)
- Co-regularized Alignment for Unsupervised Domain Adaptation (NeurIPS 2018 · similarity 0.91)
- Domain Adaptation with Asymmetrically-Relaxed Distribution Alignment (ICML 2019 · similarity 0.90)
- Domain Adaptation with Conditional Distribution Matching and Generalized Label Shift (NeurIPS 2020 · similarity 0.90)
- On Unsupervised Domain Adaptation: Pseudo Label Guided Mixup for Adversarial Prompt Tuning (AAAI 2024 · similarity 0.89)
- Toward Improving Robustness and Accuracy in Unsupervised Domain Adaptation (AAAI 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.