Research map / Label-efficient and robust learning / Recognition and long-tail
Decouple then Classify: A Dynamic Multi-view Labeling Strategy with Shared and Specific Information
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 DMVLS, an autoencoder-based semi-supervised multi-view method that decouples shared and specific information with a view discriminator and iteratively selects low-confidence samples to label (plus confident pseudo-labels).
Problem
Random sample labeling in semi-supervised multi-view classification yields unstable and unpromising accuracy, and existing methods do not decouple and jointly use shared and specific information for choosing which samples to label.
Approach
Extract shared and view-specific information with shared-parameter autoencoders, decouple them via a view discriminator, train two separate classifiers on each, and use their disagreement/low confidence to actively select samples to annotate while adding confident consistent pseudo-labels.
Key results
- At 50% labeled, DMVLS reaches 99.97 ACC on Handwritten, 97.47 on BDGP, 96.22 on Cora, 86.97 on CiteSeer, 75.63 on STL10, and 100.00 on YTB10.
- Ablation at 50% labels shows random selection drops accuracy, e.g., Cora from 96.22 to 82.01 and BDGP from 97.47 to 87.60.
- Reported std across ten random runs is smaller than most baselines, indicating more stable results.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the view discrimination loss Lvc (ablation) | DMVLS · Handwritten, BDGP, Cora, CiteSeer, STL10, YTB10 (50% labels) | ACC: 99.97±0.05 -> 99.97±0.05 (Handwritten); 97.00±0.36 -> 97.47±0.24 (BDGP); 94.83±0.50 -> 96.22±0.63 (Cora); 84.77±1.64 -> | Helped |
| Add the specific-information classification loss Lclass1 (ablation) | DMVLS · Handwritten, BDGP, Cora, CiteSeer, STL10, YTB10 (50% labels) | ACC: 99.87±0.05 -> 99.97±0.05 (Handwritten); 95.87±0.34 -> 97.47±0.24 (BDGP); 95.65±0.55 -> 96.22±0.63 (Cora); 73.62±13.01 -> | Helped |
| Add the shared-information classification loss Lclass2 (ablation) | DMVLS · Handwritten, BDGP, Cora, CiteSeer, STL10, YTB10 (50% labels) | ACC: 99.87±0.12 -> 99.97±0.05 (Handwritten); 96.96±0.34 -> 97.47±0.24 (BDGP); 95.62±0.34 -> 96.22±0.63 (Cora); 86.09±0.21 -> | Helped |
| Use uncertainty-based labeling rather than random sample selection (ablation) | DMVLS · Handwritten, BDGP, Cora, CiteSeer, STL10, YTB10 (50% labels) | ACC: 96.42±1.02 -> 99.97±0.05 (Handwritten); 87.60±0.38 -> 97.47±0.24 (BDGP); 82.01±0.55 -> 96.22±0.63 (Cora); 69.57±1.61 -> | Helped |
Surprising findings
- Merely using shared information performs better on STL10 but worse on CiteSeer than using only specific information, showing sample-dependent view informativeness.
Limitations
- Initial cold-start samples are chosen randomly, which the authors note as future work.
- Label budget is fixed per round and training stops at 50% labeled; the method is not evaluated at lower total label budgets beyond the sweep shown.
- Some baselines fail with out-of-memory on YTB10, limiting comparisons on large data.
Experimental setup
Datasets: Handwritten BDGP Cora CiteSeer STL10 YTB10
Benchmarks: Handwritten BDGP Cora CiteSeer STL10 YTB10
Metrics: ACC
Baselines: MVAR Co-GCN ERL-MVSC DSRL IMvGCN
Builds on: active learning uncertainty sampling Blum & Mitchell co-training view discriminator decoupling (Jia et al., 2020)
Closest papers by meaning
- Double-Layer Hybrid-Label Identification Feature Selection for Multi-View Multi-Label Learning (AAAI 2024 · similarity 0.88)
- Semi-Supervised Multi-View Multi-Label Learning with View-Specific Transformer and Enhanced Pseudo-Label (AAAI 2025 · similarity 0.87)
- Enhancing Multi-Label Classification via Dynamic Label-Order Learning (AAAI 2024 · similarity 0.87)
- Keep It on a Leash: Controllable Pseudo-label Generation Towards Realistic Long-Tailed Semi-Supervised Learning (NeurIPS 2025 · similarity 0.86)
- Enhancing Semi-Supervised Learning via Representative and Diverse Sample Selection (NeurIPS 2024 · similarity 0.86)
- Multi-View Multi-Label Classification via View-Label Matching Selection (AAAI 2025 · similarity 0.86)
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.