Research map / Label-efficient and robust learning / Self-supervised and contrastive
Investigating and Mitigating the Side Effects of Noisy Views for Self-Supervised Clustering Algorithms in Practical Multi-View Scenarios
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.
Investigates the noisy-view drawback in deep multi-view clustering and proposes MVCAN, which uses un-shared per-view networks, inconsistent per-view cluster assignments, and a two-level iterative optimization with view-scaling weights.
Problem
Multi-view clustering performance degrades below single-view clustering when some views are noisy, because shared parameters and forced consistent predictions let noisy views dominate the clustering objective.
Approach
Decouple network parameters per view and allow per-view soft-label matching via Hungarian-based assignment matrices, then build a robust shared learning target through a two-level iteration that scales each view's representation by its mutual information with the consensus soft labels; theoretical analysis argues this yields consistency, complementarity, and noise robustness.
Key results
- On the real-world DHA dataset MVCAN reaches 84.8% ACC versus 72.6% for the best single view DEC-BestV.
- On the noise-simulated datasets MVCAN surpasses the best comparison methods by about 7%, 5%, 1%, and 6% NMI.
- Ablation on BDGP shows removing either the un-shared parameters or the inconsistent-prediction condition drops ACC from 98.4% to 97.3% or 66.0%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add un-shared network parameters across views (condition 1) to the MVC objective (ablation) | MVCAN · BDGP and NoisyBDGP | ACC: BDGP 66.0 -> 98.4 when combined with Av; NoisyBDGP 60.2 -> 98.0 | Helped |
| Relax consistent clustering predictions across views via per-view matching Av (condition 2) (ablation) | MVCAN · BDGP and NoisyBDGP | ACC: BDGP 97.3 -> 98.4; NoisyBDGP 97.7 -> 98.0 | Helped |
| Add clustering loss L_c on top of representation loss L_r (ablation) | MVCAN · BDGP | ACC: 94.8 -> 98.4 ACC | Helped |
| Add representation loss L_r on top of clustering loss L_c (ablation) | MVCAN · BDGP | ACC: 79.8 -> 98.4 ACC | Helped |
| Replace both conditions and losses with K-means on concatenated data baseline (ablation) | MVCAN · BDGP | ACC: 64.3 vs 98.4 | Hurt |
| Use the MVC methods instead of single best view DEC on noisy views (multi-view fusion with shared parameters and consistent predictions) | existing MVC baselines · Real-world datasets (Table 2) | ACC: e.g., DMJC 64.4 vs DEC-BestV 72.6 on DHA; GP-MVC 73.1 vs 72.6; many baselines below DEC-BestV on Caltech and RGB-D | Mixed |
| Two-level iterative optimization (T1=1 removes T-level, T2=1 removes R-level) versus iterating both (ablation) | MVCAN · Multiple datasets (Figure 1) | ACC: qualitative: performance increases with T1 and T2; no numeric values in text | Helped |
| Proposed MVCAN with noise-robust multi-view objective versus best baselines on noise-simulated datasets | MVCAN · Noise-simulated datasets (Table 3) | NMI: MVCAN surpasses best comparison methods by 7%, 5%, 1%, 6% NMI | Helped |
Surprising findings
- Many existing deep MVC methods underperform the single-view DEC on the best view for real-world multi-view datasets, showing multi-view methods can be worse than single-view ones under noisy views.
Limitations
- Sensitivity to parameter initialization of the deep model remains (stated as future work).
- Sensitivity to the prior number of clusters K, which must be specified in advance.
Experimental setup
Datasets: BDGP DIGIT COIL Amazon DHA RGB-D Caltech YoutubeVideo NoisyBDGP NoisyDIGIT NoisyCOIL NoisyAmazon
Benchmarks: BDGP DIGIT COIL Amazon DHA RGB-D Caltech YoutubeVideo NoisyBDGP NoisyDIGIT NoisyCOIL NoisyAmazon
Metrics: ACC NMI
Baselines: DEC DMJC DIMC-net GP-MVC CoMVC DIMVC DSMVC DSIMVC CPSPAN SDMVC
Builds on: deep embedded clustering autoencoder representation learning K-means
Closest papers by meaning
- Partial Multi-View Clustering via Self-Supervised Network (AAAI 2024 · similarity 0.89)
- ROLL: Robust Noisy Pseudo-label Learning for Multi-View Clustering with Noisy Correspondence (CVPR 2025 · similarity 0.89)
- Robust Multi-view Clustering against Imperfect Information (NeurIPS 2026 · similarity 0.88)
- Self-supervised Trusted Contrastive Multi-view Clustering with Uncertainty Refined (AAAI 2025 · similarity 0.88)
- Automatically Identify and Rectify: Robust Deep Contrastive Multi-view Clustering in Noisy Scenarios (ICML 2025 · similarity 0.88)
- Mixture of Experts as Representation Learner for Deep Multi-View Clustering (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.