Research map / Label-efficient and robust learning / Semi-supervised learning
Enhancing Semi-supervised Domain Adaptation via Effective Target Labeling
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 EFTL, an active-learning plus pseudo-labeling framework that automatically selects informative target samples for labeling in semi-supervised domain adaptation, and couples it with MME and FixMatch+MME.
Problem
Semi-supervised domain adaptation (SSDA) assumes an equal number of labeled target samples per class, which requires heavy manual annotation and may select uninformative or outlier samples. The paper targets automatic selection of a fixed annotation budget of informative target samples.
Approach
Non-maximal degree node suppression (NDNS) builds a directed kNN graph over target features (accepted/acceptive neighbors), iteratively queries max out-degree nodes and suppresses their 2-hop neighbors, weighted by model uncertainty; clustering consistency (GMM and K-means agreement) pseudo-labels low-confidence target samples.
Key results
- EFTL improves MME by 2.6 (1-shot) and 3.7 (3-shot) average points on Office-Home and by over 4% on DomainNet.
- FixMME+EFTL reaches 74.4% (1-shot) and 78.7% (3-shot) average accuracy on Office-Home, outperforming FixMME by about 2.7 and 3.5 points.
- Existing ADA methods TQS and SDM-A combined with SSDA models often underperform the SSDA-only baseline (e.g., MME 70.3 vs TQS+MME), while EFTL improves it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply EFTL target labeling with MME | MME · DomainNet | accuracy: +4.1 (1-shot); +4.4 (3-shot) | Helped |
| Apply EFTL target labeling with FixMME | FixMME · DomainNet | accuracy: +3.2 (1-shot); +3.2 (3-shot) | Helped |
| Apply EFTL target labeling with MME | MME · Office-Home | accuracy: +2.6 (1-shot); +3.7 (3-shot) | Helped |
| Apply EFTL target labeling with FixMME | FixMME · Office-Home | accuracy: +2.7 (1-shot); +3.5 (3-shot) | Helped |
| Add NDNS target selection (ablation) | FixMME · Office-Home A→C | accuracy: 64.0 -> 64.0 (1-shot); 65.8 -> 69.9 (3-shot) | Mixed |
| Add NDNS target selection (ablation) | FixMME · Office-Home C→P | accuracy: 75.2 -> 78.7 (1-shot); 81.4 -> 84.6 (3-shot) | Helped |
| Add NDNS target selection (ablation) | MME · Office-Home A→C | accuracy: 59.6 -> 62.9 (1-shot); 63.6 -> 67.9 (3-shot) | Helped |
| Add NDNS target selection (ablation) | MME · Office-Home C→P | accuracy: 74.5 -> 78.6 (1-shot); 79.3 -> 82.1 (3-shot) | Helped |
| Add clustering consistency pseudo-labeling (CC) (ablation) | NDNS + FixMME · Office-Home A→C | accuracy: 64.0 -> 65.7 (1-shot); 69.9 -> 70.3 (3-shot) | Helped |
| Add clustering consistency pseudo-labeling (CC) (ablation) | NDNS + FixMME · Office-Home C→P | accuracy: 78.7 -> 79.6 (1-shot); 84.6 -> 84.6 (3-shot) | Mixed |
| Add clustering consistency pseudo-labeling (CC) (ablation) | NDNS + MME · Office-Home A→C | accuracy: 62.9 -> 63.1 (1-shot); 67.9 -> 68.2 (3-shot) | Helped |
Surprising findings
- Existing active domain adaptation methods (TQS, SDM-A) can deteriorate SSDA performance when combined with MME, sometimes below the SSDA-only setting.
Limitations
- Pseudo-labeling and active query rely on a source-dominated model whose features may be unreliable under large domain shift
- Hyperparameters such as M1, M2 and alpha depend on dataset scale and were tuned on specific transfer tasks
Experimental setup
Datasets: Office-31 Office-Home DomainNet
Benchmarks: Office-31 Office-Home DomainNet
Metrics: classification accuracy
Baselines: S+T ENT CDAC DECOTA MME+SLA MCL TQS SDM-A
Builds on: MME FixMatch non-maximum suppression Yang et al. 2021b neighbor definitions
Closest papers by meaning
- Universal Semi-Supervised Domain Adaptation by Mitigating Common-Class Bias (CVPR 2024 · similarity 0.90)
- Diffusion-Driven Progressive Target Manipulation for Source-Free Domain Adaptation (NeurIPS 2025 · similarity 0.88)
- Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.88)
- A Unified Framework for Heterogeneous Semi-supervised Learning (CVPR 2025 · similarity 0.87)
- CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 2025 · similarity 0.87)
- Gradient-Based Sample Selection for Black-Box Universal Domain Adaptation (AAAI 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.