Research map / Label-efficient and robust learning / Domain adaptation and generalization
Collaborative Semantic Consistency Alignment for Blended-Target 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 CSCA, a blended-target domain adaptation method combining sliced Wasserstein distribution alignment, graph-network semantic propagation across source and all sub-target domains, and a double (prediction and centroid) consistency regularization.
Problem
Blended-target domain adaptation, where a source model must adapt to a mixture of unlabeled sub-target domains with distinct statistics, while exploiting semantic correlations among sub-targets and reducing shifts among them.
Approach
Aligns global source-blended-target feature distributions with sliced Wasserstein distance, uses a graph network trained on pseudo-label-derived affinity matrices to share semantic information across all domains, and applies consistency regularization over both category assignments and centroid assignments between weak and strong views to suppress domain-specific information.
Key results
- On Office-Home, CSCA reaches 73.0% average accuracy, versus 71.9% for MTDA-PDT and 71.1% for MCDA (the best prior baseline).
- On ImageCLEF-DA, CSCA reaches 82.8% average accuracy, versus 82.7% for HTA.
- On Office-31, CSCA reaches 91.0% average accuracy, versus 89.6% for MCDA; on task A→W/D it exceeds MCDA by 3.4%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CSCA instead of DCGCT | BTDA with ResNet-50 · Office-31 | accuracy (%): 88.8 -> 91.0 | Helped |
| use CSCA instead of MTDA-PDT | BTDA with ResNet-50 · Office-Home | accuracy (%): 71.9 -> 73.0 | Helped |
| use CSCA instead of HTA | BTDA with ResNet-50 · ImageCLEF-DA | accuracy (%): 82.7 -> 82.8 | Helped |
| add SWD loss (ablation) | CSCA · Office-Home | accuracy (%): 70.7 -> 73.0 | Helped |
| add graph loss (ablation) | CSCA · Office-Home | accuracy (%): 66.4 -> 73.0 | Helped |
| add consistency loss (ablation) | CSCA · Office-Home | accuracy (%): 69.9 -> 73.0 | Helped |
| add feature consistency to prediction consistency (ablation) | CSCA · Office-31 | accuracy (%): 89.6 -> 91.0 | Helped |
| add prediction consistency to feature consistency (ablation) | CSCA · Office-31 | accuracy (%): 90.1 -> 91.0 | Helped |
Limitations
- Pseudo-label quality is critical; the graph and consistency components depend on threshold-based pseudo-labels.
- Evaluation limited to image classification benchmarks with a single ResNet-50 backbone.
Experimental setup
Datasets: Office-31 Office-Home ImageCLEF-DA DomainNet
Benchmarks: Office-31 Office-Home ImageCLEF-DA
Metrics: classification accuracy (mean over sub-target domains)
Baselines: Source only DAN DANN CDAN JAN AMEAN CGCT MT-MTDA DCGCT MTDA-PDT MCDA AMDA
Builds on: sliced Wasserstein distance (Lee et al. 2019) AMEAN CGCT MCDA
Closest papers by meaning
- Invertible Projection and Conditional Alignment for Multi-Source Blended-Target Domain Adaptation (AAAI 2025 · similarity 0.89)
- Class Overwhelms: Mutual Conditional Blended-Target Domain Adaptation (AAAI 2023 · similarity 0.88)
- Multi-Target Domain Adaptation With Collaborative Consistency Learning (CVPR 2021 · similarity 0.88)
- Open-Set Graph Domain Adaptation via Separate Domain Alignment (AAAI 2024 · similarity 0.88)
- Target Semantics Clustering via Text Representations for Robust Universal Domain Adaptation (AAAI 2025 · similarity 0.87)
- Pairwise Alignment Improves Graph Domain Adaptation (ICML 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.