Research map / Label-efficient and robust learning / Recognition and long-tail
CFDM: Contrastive Fusion and Disambiguation for Multi-View Partial-Label Learning
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 CFDM, a multi-view partial-label learning framework that fuses view-specific autoencoder features via inter-/intra-view contrastive learning and disambiguates candidate labels with contrastive class prototypes and pseudo-label updates.
Problem
Multi-view partial-label learning (MVPLL), where each instance has multiple view features and a candidate label set containing exactly one correct label; prior methods handle only two views and do not link multi-view features to ambiguous labels directly.
Approach
Mutually reinforcing multi-view contrastive fusion (inter-view and classifier-guided intra-view contrast over view-specific autoencoder representations) and multi-class contrastive prototype disambiguation that maintains per-view class prototypes and updates pseudo labels via nearest-prototype assignment, scaling to any number of views.
Key results
- CFDM achieves the best accuracy on MSRCv1, Caltech101-7, Mfeat, CCV and Caltech101-all across the six configurations, e.g., 0.668±0.011 on Caltech101-all (r=1, p=0.3) vs. 0.653 for UCL.
- On Scene15 CFDM is superior to other methods in over 94.4% of cases.
- Ablation on Caltech101-all shows removing any of the reconstruction, inter-view, intra-view, or prototype losses lowers accuracy (0.641-0.656 vs. 0.668).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CFDM instead of LMVCAT | MVPLL · MSRCv1 (r=1, p=0.3) | accuracy: 0.966±0.036 -> 0.971±0.021 | Helped |
| Use CFDM instead of UCL | MVPLL · Caltech101-7 (r=1, p=0.3) | accuracy: 0.981±0.013 -> 0.982±0.002 | Helped |
| Use CFDM instead of GLADE | MVPLL · Mfeat (r=1, p=0.3) | accuracy: 0.979±0.012 -> 0.987±0.005 | Helped |
| Use CFDM instead of UCL | MVPLL · Scene15 (r=1, p=0.3) | accuracy: 0.762±0.012 -> 0.778±0.010 | Helped |
| Use CFDM instead of UCL | MVPLL · CCV (r=1, p=0.3) | accuracy: 0.534±0.012 -> 0.537±0.011 | Helped |
| Use CFDM instead of UCL | MVPLL · Caltech101-all (r=1, p=0.3) | accuracy: 0.653±0.008 -> 0.668±0.011 | Helped |
| Add reconstruction loss Lrec (ablation) | CFDM · Caltech101-all (r=1, p=0.3) | accuracy: 0.656±0.021 -> 0.668±0.011 | Helped |
| Add inter-view contrastive loss Lcont1 (ablation) | CFDM · Caltech101-all (r=1, p=0.3) | accuracy: 0.641±0.010 -> 0.668±0.011 | Helped |
| Add intra-view contrastive loss Lcont2 (ablation) | CFDM · Caltech101-all (r=1, p=0.3) | accuracy: 0.642±0.015 -> 0.668±0.011 | Helped |
| Add prototype contrastive loss Lcont3 (ablation) | CFDM · Caltech101-all (r=1, p=0.3) | accuracy: 0.655±0.013 -> 0.668±0.011 | Helped |
Limitations
- Evaluated only on synthetically generated partial-label versions of existing multi-view datasets.
- Reports only test accuracy with 5-fold cross-validation; many hyperparameters (lambdas, learning rate) are tuned per setting.
Experimental setup
Datasets: MSRCv1 Caltech101-7 Mfeat Scene15 CCV Caltech101-all
Benchmarks: MSRCv1 Caltech101-7 Mfeat Scene15 CCV Caltech101-all
Metrics: test accuracy
Baselines: GRADIS GLADE ML-BVAE LMVCAT UCL TERIAL
Builds on: partial-label learning multi-view contrastive clustering view-specific autoencoders
Closest papers by meaning
- Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding (CVPR 2025 · similarity 0.89)
- Partial Multi-View Multi-Label Classification via Semantic Invariance Learning and Prototype Modeling (ICML 2024 · similarity 0.89)
- Learning Compact Semantic Information for Incomplete Multi-View Missing Multi-Label Classification (ICML 2025 · similarity 0.89)
- Double-Layer Hybrid-Label Identification Feature Selection for Multi-View Multi-Label Learning (AAAI 2024 · similarity 0.89)
- Semi-Supervised Multi-View Multi-Label Learning with View-Specific Transformer and Enhanced Pseudo-Label (AAAI 2025 · similarity 0.89)
- Multi-View Multi-Label Classification via View-Label Matching Selection (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.