Research map / Learning theory and algorithms / Clustering
Robust Contrastive Multi-view Clustering against Dual Noisy Correspondence
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 CANDY, a plug-and-play robust contrastive multi-view clustering method that handles dual noisy correspondence (false positives and false negatives) via context-based semantic mining and spectral denoising of a high-order affinity graph.
Problem
Contrastive multi-view clustering assumes faultless cross-view correspondence, but noise corrupts both positive pairs (false positives) and negative pairs (false negatives, semantically associated samples sampled as negatives), a setting called Dual Noisy Correspondence (DNC).
Approach
Treats each point's inter-view affinity vector as a context embedding to build a high-order affinity graph that recovers false negatives globally, then applies SVD-based spectral truncation to denoise the graph and use it as a soft pseudo target for a contrastive loss.
Key results
- Average ACC over five datasets at 0% FP ratio: CANDY 51.9 vs. DIVIDE 49.6 (ACC) and 35.0 vs. 29.7 (ARI).
- At 50% FP ratio, CANDY averages 47.5 ACC vs. DIVIDE 38.0 ACC; on Caltech-101 at 20% FP, CANDY reaches 65.9 ACC vs. DIVIDE 48.3.
- On incomplete MvC benchmarks (50% missing views), CANDY averages 48.1 ACC, close to DIVIDE's 48.7.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace DIVIDE's loss with CANDY's DNC-robust target and loss | DIVIDE · Average across five datasets, FP ratio 20% | ACC: 44.4 -> 50.3 | Helped |
| replace DIVIDE's loss with CANDY's DNC-robust target and loss | DIVIDE · Average across five datasets, FP ratio 50% | ACC: 38.0 -> 47.5 | Helped |
| replace DIVIDE's loss with CANDY's DNC-robust target and loss | DIVIDE · Average across five datasets, FP ratio 80% | ARI: 18.2 -> 27.1 | Helped |
| add re-alignment (ablation) | CANDY · Caltech-101, FP ratio 20% | ACC: 46.9 -> 49.9 | Helped |
| add re-alignment (ablation) | CANDY · NUS-WIDE, FP ratio 20% | ACC: 57.5 -> 58.6 | Helped |
| add re-alignment (ablation) | CANDY · Caltech-101, FP ratio 50% | ACC: 35.6 -> 41.1 | Helped |
| add re-alignment (ablation) | CANDY · NUS-WIDE, FP ratio 50% | ACC: 44.2 -> 46.6 | Helped |
| add spectral-based correspondence denoising (SCD) (ablation) | CANDY · Caltech-101, FP ratio 20% | ACC: 49.9 -> 56.5 | Helped |
| add spectral-based correspondence denoising (SCD) (ablation) | CANDY · NUS-WIDE, FP ratio 20% | ACC: 58.6 -> 58.1 | Hurt |
| add spectral-based correspondence denoising (SCD) (ablation) | CANDY · Caltech-101, FP ratio 50% | ACC: 41.1 -> 54.0 | Helped |
| add spectral-based correspondence denoising (SCD) (ablation) | CANDY · NUS-WIDE, FP ratio 50% | ACC: 46.6 -> 55.6 | Helped |
| add context-based semantic mining (CSM) and its resulting pseudo target (ablation) | CANDY · Caltech-101, FP ratio 20% | ACC: 56.5 -> 65.0 | Helped |
Surprising findings
- Applying the noisy correspondence learning method RCL on top of DIVIDE degrades substantially under DNC, e.g., Caltech-101 ACC falls from 44.9 to 16.7 as FP ratio rises from 0 to 0.8.
- At 0% FP ratio on Scene-15, CANDY (42.0 ACC) underperforms DIVIDE (49.1 ACC), so the average gains are not uniform across datasets.
Limitations
- Gains are not uniform: on Scene-15 at 0% FP CANDY is worse than DIVIDE, and on NUS-WIDE at 80% FP it trails DIVIDE in some metrics.
- Denoising hyperparameter eta and lambda are fixed at 0.2 without tuning; sensitivity is only shown on two datasets.
- Does not address simultaneous noisy correspondence and missing modalities, which the authors leave for future work.
Experimental setup
Datasets: Scene-15 Caltech-101 LandUse-21 Reuters NUS-WIDE
Benchmarks: Scene-15 Caltech-101 LandUse-21 Reuters NUS-WIDE
Metrics: ACC NMI ARI
Baselines: DCCAE BMVC PVC MVCLN SURE GCFAgg CGCN DIVIDE RCL PMVC DAIMC EERIMVC
Builds on: DIVIDE contrastive learning partially view-aligned learning (PVP) AECoKM
Closest papers by meaning
- Decoupled Contrastive Multi-View Clustering with High-Order Random Walks (AAAI 2024 · similarity 0.89)
- Robust Multi-view Clustering against Imperfect Information (NeurIPS 2026 · similarity 0.88)
- Incomplete Multi-view Clustering via Diffusion Contrastive Generation (AAAI 2025 · similarity 0.87)
- EASEMVC:Efficient Dual Selection Mechanism for Deep Multi-View Clustering (CVPR 2025 · similarity 0.87)
- Medusa: A Multi-Scale High-order Contrastive Dual-Diffusion Approach for Multi-View Clustering (CVPR 2025 · similarity 0.87)
- ROLL: Robust Noisy Pseudo-label Learning for Multi-View Clustering with Noisy Correspondence (CVPR 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.