Research map / Label-efficient and robust learning / Self-supervised and contrastive
Self-supervised Trusted Contrastive Multi-view Clustering with Uncertainty Refined
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 STCMC-UR, a trusted contrastive multi-view clustering method that fuses two evidence-generation BBAs and multiple views via Dempster-Shafer theory and uses uncertainty-filtered pseudo-labels to remove false negatives in contrastive learning.
Problem
Contrastive multi-view clustering ignores clustering reliability and false negative pairs in negative sampling, and lacks uncertainty estimation, limiting use in safety-critical settings.
Approach
Combines evidential deep learning (Dirichlet/Dempster-Shafer fusion of two BBAs per view and across views) with feature-level contrastive learning, where only high-confidence (low-uncertainty) pseudo-labels are used to exclude false negatives and view weights are derived from clustering results.
Key results
- On Caltech-2V, ACC 71.9 and NMI 66.1, versus second-best SPDMC at 64.4/50.6.
- On COIL20 and NUS22 the method reports ACC of 100 and NMI of 100 on COIL20 (as stated in the table), and 18.5 ACC on NUS22 vs 19.1 for DIVIDE.
- Ablation: removing LCL or LCE lowers performance on all datasets, e.g. Caltech-2V ACC drops from 71.9 to 63.6 without LCL and to 66.4 without LCE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add uncertainty-refined contrastive learning (LCL) and view fusion consistency loss (LCE) to LC (ablation) | STCMC-UR · Caltech-2V | ACC: 48.9 -> 71.9 | Helped |
| Add uncertainty-refined contrastive learning (LCL) and view fusion consistency loss (LCE) to LC (ablation) | STCMC-UR · Caltech-3V | ACC: 23.4 -> 76.3 | Helped |
| Add uncertainty-refined contrastive learning (LCL) and view fusion consistency loss (LCE) to LC (ablation) | STCMC-UR · COIL20 | ACC: 86.7 -> 100 | Helped |
| Add uncertainty-refined contrastive learning (LCL) and view fusion consistency loss (LCE) to LC (ablation) | STCMC-UR · Event8 | ACC: 43.0 -> 56.9 | Helped |
| Add uncertainty-refined contrastive learning (LCL) and view fusion consistency loss (LCE) to LC (ablation) | STCMC-UR · NUS22 | ACC: 11.0 -> 18.5 | Helped |
| Add LCL to LC (ablation) | STCMC-UR · Caltech-2V | ACC: 48.9 -> 63.6 | Helped |
| Add LCE to LC (ablation) | STCMC-UR · Caltech-2V | ACC: 48.9 -> 66.4 | Helped |
| Remove LCE from the full method (ablation) | STCMC-UR · Caltech-3V | ACC: 76.3 -> 74.2 | Hurt |
| Remove LCL from the full method (ablation) | STCMC-UR · Caltech-3V | ACC: 76.3 -> 27.9 | Hurt |
| Add LCL to LC (ablation) | STCMC-UR · COIL20 | ACC: 86.7 -> 91.0 | Helped |
| Add LCE to LC (ablation) | STCMC-UR · COIL20 | ACC: 86.7 -> 90.0 | Helped |
| Remove LCE from the full method (ablation) | STCMC-UR · NUS22 | ACC: 18.5 -> 16.7 | Hurt |
Surprising findings
- Table 2 reports 100% ACC and NMI on COIL20 while the text states COIL20 results are 56.9/40.1 in the table, an apparent inconsistency in the reported numbers (Table 2 column labels appear shifted).
Limitations
- Does not handle incomplete multi-view data where views have missing values.
- Fixed uncertainty threshold epsilon uniformly applied across datasets.
Experimental setup
Datasets: Caltech-2V Caltech-3V COIL20 Event8 NUS22
Benchmarks: Caltech-2V Caltech-3V COIL20 Event8 NUS22
Metrics: ACC NMI
Baselines: KM Ncuts MvSCN EAMC MVC-VAE DEMVC SiMVC CoMVC MFLVC SPDMC ICMVC DIVIDE
Builds on: contrastive multi-view clustering evidential deep learning Dempster-Shafer theory NT-Xent
Closest papers by meaning
- Partial Multi-View Clustering via Self-Supervised Network (AAAI 2024 · similarity 0.89)
- Investigating and Mitigating the Side Effects of Noisy Views for Self-Supervised Clustering Algorithms in Practical Multi-View Scenarios (CVPR 2024 · similarity 0.88)
- Decoupled Contrastive Multi-View Clustering with High-Order Random Walks (AAAI 2024 · similarity 0.88)
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.87)
- EASEMVC:Efficient Dual Selection Mechanism for Deep Multi-View Clustering (CVPR 2025 · similarity 0.87)
- Ambiguous Instance-Aware Contrastive Network with Multi-Level Matching for Multi-View Document 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.