Research map / Label-efficient and robust learning / Semi-supervised learning
Semi-Supervised Multi-View Multi-Label Learning with View-Specific Transformer and Enhanced Pseudo-Label
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 SMVTEP, a semi-supervised multi-view multi-label method combining GAN-based shared/specific representation learning, view-specific classifiers with manifold constraints, a label-guided transformer fusion, and category-enhanced pseudo-labeling with Mixup.
Problem
Multi-view multi-label classification when most instances are unlabeled and labeled data is scarce, requiring richer view semantics, refined classification under limited labels, and higher-quality supervision from pseudo-labels.
Approach
Integrates GAN-based shared/specific feature extraction (adversarial view-consistency plus information-theoretic contrastive loss) with a label-guided transformer that injects confidence scores and class tokens into self-attention, and a class-distribution-aware pseudo-labeling strategy after Mixup augmentation.
Key results
- On Corel 5k at LER 6%, SMVTEP reaches AP 32.52% versus 26.97% for the best baseline (LMVCAT), and 1-RL 84.80 vs 81.28.
- On ESP Game at LER 21%, SMVTEP AP is 29.72% vs 27.43% for LMVCAT; at LER 3% it is 25.05% vs 20.05%.
- Ablation on Corel 5k (AP): removing the label-guided transformer drops AP from 32.09 to 29.97 in the shown configuration, and removing the pseudo-label strategy drops it to 20.59.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SMVTEP instead of SCTML | Yeast · Yeast | AP: 70.65±1.44 -> 72.95±1.21 | Helped |
| use SMVTEP instead of LMVCAT | Corel 5k · Corel 5k | AP: 26.97±1.09 -> 32.52±0.54 | Helped |
| use SMVTEP instead of LMVCAT | VOC 2007 · VOC 2007 | AP: 49.35±1.47 -> 54.18±1.03 | Helped |
| use SMVTEP instead of DICNet | ESP Game · ESP Game | AP: 22.68±0.24 -> 27.22±0.63 | Helped |
| use SMVTEP instead of LMVCAT | IAPR TC-12 · IAPR TC-12 | AP: 22.99±0.44 -> 27.76±0.21 | Helped |
| use SMVTEP instead of LMVCAT | MIR FLICKR · MIR FLICKR | AP: 58.19±1.10 -> 60.27±0.25 | Helped |
| add Generative Adversarial Networks for feature extraction (ablation) | SMVTEP · Yeast | AP: 71.93 -> 75.51 | Helped |
| add Generative Adversarial Networks for feature extraction (ablation) | SMVTEP · Corel 5k | AP: 31.34 -> 32.09 | Helped |
| add Generative Adversarial Networks for feature extraction (ablation) | SMVTEP · VOC 2007 | AP: 51.65 -> 54.41 | Helped |
| add label-guided transformer fusion (ablation) | SMVTEP · Yeast | AP: 72.05 -> 75.51 | Helped |
| add label-guided transformer fusion (ablation) | SMVTEP · Corel 5k | AP: 29.97 -> 32.09 | Helped |
| add label-guided transformer fusion (ablation) | SMVTEP · VOC 2007 | AP: 50.77 -> 54.41 | Helped |
Surprising findings
- Single-view semi-supervised baselines (ESMC, SCTML) run per view and perform worse than multi-view incomplete-label methods on most datasets.
Limitations
- Future work acknowledged: label noise and class imbalance are not addressed.
- Hyperparameters (margin, alpha, eta thresholds) are fixed by manual choice; sensitivity studies are limited to lambda1 and lambda2.
Experimental setup
Datasets: Yeast Corel 5k VOC 2007 ESP Game IAPR TC-12 MIR FLICKR
Benchmarks: Yeast Corel 5k VOC 2007 ESP Game IAPR TC-12 MIR FLICKR
Metrics: 1-HL 1-RL AP AUC
Baselines: iMVWL TM3L NAIM3L DICNet DD-IMvMLC LMVCAT ESMC SCTML
Builds on: Class-distribution-aware pseudo-labeling Multi-view multi-label learning with view-specific information extraction Mixup
Closest papers by meaning
- View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning (CVPR 2024 · similarity 0.90)
- Double-Layer Hybrid-Label Identification Feature Selection for Multi-View Multi-Label Learning (AAAI 2024 · similarity 0.90)
- Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding (CVPR 2025 · similarity 0.89)
- CFDM: Contrastive Fusion and Disambiguation for Multi-View Partial-Label Learning (AAAI 2025 · similarity 0.89)
- Multi-View Multi-Label Classification via View-Label Matching Selection (AAAI 2025 · similarity 0.88)
- Partial Multi-View Multi-Label Classification via Semantic Invariance Learning and Prototype Modeling (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.