Research map / Learning theory and algorithms / Clustering
Sample-Level Cross-View Similarity Learning for Incomplete Multi-View Clustering
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 SCSL, a sample-level cross-view similarity learning method for incomplete multi-view clustering that builds a complete similarity matrix from inter- and intra-view sample pairs, with shared projections, consensus representation and graph Laplacian regularization.
Problem
Incomplete multi-view clustering where similarity-based methods fail to produce a complete unified similarity matrix because some sample pairs never co-occur in a view, and column-sum constraints cause abnormal similarity values under missingness.
Approach
Constructs the consensus similarity from all cross-view sample pairs (not only within-view pairs) by projecting every view into a common k-dimensional space with orthogonal projections, adds a balancing weight per pair, and couples the similarity with a consensus representation via graph Laplacian regularization.
Key results
- On MSRCV, SCSL reaches 79.21% ACC versus 74.72% for the second-best AGC-IMC, a gain of 6.01 points over the second-best method.
- SCSL outperforms the second-best method in ACC by 4.61% on ORL, 4.79% on ProteinFold, 0.92% on Wiki, 3.63% on CCV and 3.41% on SUNRGBD.
- Removing the Laplacian regularization term lowers ACC on MSRCV from 79.21% to 66.43% (with large variance), showing the term's contribution.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SCSL instead of the strongest reported baseline | Incomplete multi-view clustering · MSRCV, ORL, ProteinFold, Wiki, CCV, SUNRGBD | ACC: +6.01%, 4.61%, 4.79%, 0.92%, 3.63%, and 3.41%, respectively | Helped |
| add graph Laplacian regularization (ablation) | SCSL · MSRCV | ACC: 66.43±56.21 -> 79.21±0.78 | Helped |
| add graph Laplacian regularization (ablation) | SCSL · ORL, ProteinFold, Wiki, CCV, SUNRGBD | ACC | Helped |
Surprising findings
- Without the Laplacian term, results have very large standard deviations (e.g., ±56.21 on MSRCV ACC), indicating instability.
Limitations
- Hyperparameter search grid is coarse (beta in {0.001, 1, 10}, lambda in {0.001, 0.1, 1}).
- Evaluated only on six datasets with randomly generated missingness.
Experimental setup
Datasets: MSRCV ORL ProteinFold Wiki CCV SUNRGBD
Benchmarks: MSRCV ORL ProteinFold Wiki CCV SUNRGBD
Metrics: ACC NMI Purity
Baselines: BSV MIC MKKM-IK DAIMC UEAF MKKM-MKC V3H AGC-IMC HCP-IMSC
Builds on: similarity-based incomplete multi-view clustering graph Laplacian regularization generalized power iteration
Closest papers by meaning
- Collaborative Similarity Fusion and Consistency Recovery for Incomplete Multi-view Clustering (AAAI 2025 · similarity 0.92)
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.91)
- Incomplete and Unpaired Multi-View Graph Clustering with Cross-View Feature Fusion (AAAI 2025 · similarity 0.91)
- Fast Incomplete Multi-view Clustering by Flexible Anchor Learning (ICML 2025 · similarity 0.91)
- Fast Incomplete Multi-view Clustering with Adaptive Similarity Completion and Reconstruction (AAAI 2025 · similarity 0.90)
- Towards Resource-friendly, Extensible and Stable Incomplete Multi-view Clustering (ICML 2024 · similarity 0.90)
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.