Research map / Label-efficient and robust learning / Self-supervised and contrastive
Partial Multi-View Clustering via Self-Supervised Network
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 PVC-SSN, a deep partial multi-view clustering network that combines multi-view supervised contrastive learning (with pseudo-labels from self-expression spectral clustering), a self-expression layer, and a multi-view decoder, evaluated on BDGP, MNIST, and HW.
Problem
Partial multi-view clustering, where samples are missing in one or more views, struggles to learn a consistent cluster structure across views with large differences and ignores discriminative category information in the unsupervised setting.
Approach
Uses cluster pseudo-labels derived from a shared self-expression coefficient matrix to define positive/negative pairs for multi-view contrastive learning, jointly trained with self-expression and reconstruction losses so that the subspace becomes more cluster-friendly.
Key results
- On BDGP at 0.9 missing ratio, PVC-SSN reaches 60.32% ACC versus 59.01% for iCmSC; at 0.7 it reaches 80.36% vs 74.77%.
- On MNIST (four views), PVC-SSN is reported as 2-7% ACC above iCmSC across missing ratios, e.g. 57.75% vs 50.89% at 0.9 missing rate.
- On HW (three views), PVC-SSN attains 82.45% ACC at 0.9 missing rate, above iCmSC's 76.10%, but is slightly below iCmSC at 0.1 (94.85 vs 95.00).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PVC-SSN instead of the strongest listed baseline, iCmSC | Partial multi-view clustering · BDGP, missing rates 0.9, 0.7, 0.5, 0.3, 0.1 | ACC (%): 0.6032±0.0144 vs 0.5901±0.0079; 0.8036±0.0109 vs 0.7477±0.0043; 0.9156±0.0087 vs 0.8845±0.0030; 0.9304±0.0073 vs 0.9210± | Helped |
| Use PVC-SSN instead of the strongest listed baseline, iCmSC | Partial multi-view clustering · MNIST, missing rates 0.9, 0.7, 0.5, 0.3, 0.1 | ACC (%): 0.5775±0.0094 vs 0.5089±0.0074; 0.5835±0.0073 vs 0.5665±0.0163; 0.6277±0.0104 vs 0.5834±0.0089; 0.6335±0.0098 vs 0.6012± | Helped |
| Use PVC-SSN instead of the strongest listed baseline | Partial multi-view clustering · HW, missing rates 0.9, 0.7, 0.5, 0.3, 0.1 | ACC (%): 0.8245±0.0057 vs 0.7610±0.0062; 0.8835±0.0132 vs 0.8517±0.0177; 0.9415±0.0046 vs 0.9158±0.0085; 0.9520±0.0066 vs 0.9450± | Mixed |
| Add the multi-view contrastive module (ablation) | PVC · BDGP, missing rates 0.9, 0.7, 0.5, 0.3, 0.1 | ACC (%): 0.5256 -> 0.6032; 0.7016 -> 0.8036; 0.8272 -> 0.9156; 0.8912 -> 0.9304; 0.9404 -> 0.9616 | Helped |
Surprising findings
- On HW at 0.1 missing ratio, PVC-SSN (94.85%) is slightly below iCmSC (95.00%) and PVC-GAN is close, indicating it is not uniformly best.
- The reported MNIST ACC claim of 2-7% over iCmSC is inconsistent with the table, where at 0.1 missing ratio the gap is about 0.5 points (ACC 66.2 vs 63.2).
Limitations
- Experiments are limited to three small benchmark datasets (BDGP, MNIST subset, HW).
- Performance on HW at low missing ratios is slightly below iCmSC.
- Paper reports fluctuations and sensitivity to lambda hyperparameters with a per-dataset best setting.
Experimental setup
Datasets: BDGP MNIST HW
Benchmarks: BDGP MNIST HW
Metrics: ACC NMI running time
Baselines: SC AMGL RMSC ConSC IMG GPVC PVC-GAN iCmSC PVC
Builds on: subspace clustering supervised contrastive learning PVC-GAN
Closest papers by meaning
- Self-supervised Trusted Contrastive Multi-view Clustering with Uncertainty Refined (AAAI 2025 · similarity 0.89)
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (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 0.89)
- Towards Resource-friendly, Extensible and Stable Incomplete Multi-view Clustering (ICML 2024 · similarity 0.89)
- Contrastive Multi-view Subspace Clustering via Tensor Transformers Autoencoder (AAAI 2025 · similarity 0.88)
- Incomplete and Unpaired Multi-View Graph Clustering with Cross-View Feature Fusion (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.