Research map / Learning theory and algorithms / Clustering
Learning Cluster-Wise Anchors for 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 CAMVC, an anchor-based multi-view clustering method that learns cluster-wise anchors guided by a prior cluster indicator matrix and orthogonal cluster centroids, solved by alternating closed-form updates.
Problem
Anchor-based multi-view clustering typically enforces orthogonality among anchors, which can yield imbalanced, non-representative anchors that ignore semantic cluster structure, limiting subspace representation discrimination for large-scale clustering.
Approach
Assume latent anchors share the cluster structure of the data via a predefined prior cluster indicator Y; anchors are pulled toward orthogonal cluster centroids H_i Y, so within-cluster anchors are semantically correlated and between-cluster anchors are exclusive, and this is jointly learned with the consensus representation Z.
Key results
- CAMVC achieves average ACC score 0.8049 and average rank 2.00 across seven datasets, versus 0.6928 for the best baseline AWMVC (rank 4.00).
- On BBCSport, CAMVC reaches ACC 0.8997 and NMI 0.8951, far above the best baseline SFMC (ACC 0.8254).
- Ablation removing the cluster-structure term with orthogonal anchors (Ours-a) lowers ACC on MSRC from 0.8418 to 0.7507 and on Caltech101 from 0.6301 to 0.5597.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add cluster-structure regularization and use cluster-wise anchors instead of the Ours-a anchor variant (ablation) | CAMVC · MSRC | ACC: 0.7507 -> 0.8418 | Helped |
| Add cluster-structure regularization and use cluster-wise anchors instead of the Ours-a anchor variant (ablation) | CAMVC · BBCSport | ACC: 0.8956 -> 0.8997 | Helped |
| Add cluster-structure regularization and use cluster-wise anchors instead of the Ours-a anchor variant (ablation) | CAMVC · Wiki | ACC: 0.5201 -> 0.5386 | Helped |
| Add cluster-structure regularization and use cluster-wise anchors instead of the Ours-a anchor variant (ablation) | CAMVC · Notting-Hill | ACC: 0.9254 -> 0.9521 | Helped |
| Add cluster-structure regularization and use cluster-wise anchors instead of the Ours-a anchor variant (ablation) | CAMVC · Caltech101 | ACC: 0.5597 -> 0.6301 | Helped |
| Add cluster-structure regularization and use cluster-wise anchors instead of the Ours-a anchor variant (ablation) | CAMVC · Fashion | ACC: 0.7384 -> 0.7897 | Helped |
| Add cluster-structure regularization and use cluster-wise anchors instead of the Ours-a anchor variant (ablation) | CAMVC · MNIST | ACC: 0.9808 -> 0.9826 | Helped |
| Increase the number of anchors per cluster from one to multiple anchors (ablation) | CAMVC · datasets shown in Figure 3 | ACC | Helped |
Surprising findings
- Orthogonal-anchor variant performs notably worse than the cluster-wise variant, suggesting orthogonality may cause imbalanced anchors.
Limitations
- Prior cluster indicator Y is pre-defined with fixed m anchors per cluster, limiting flexibility.
- Objective is non-convex; only convergence to a local minimum is shown.
- Matrix-factorization baselines (OPMC, AWMVC) are faster than CAMVC.
Experimental setup
Datasets: MSRC BBCSport Wiki Notting-Hill Caltech101 Fashion MNIST
Benchmarks: MSRC BBCSport Wiki Notting-Hill Caltech101 Fashion MNIST
Metrics: ACC NMI Purity Fscore
Baselines: LMVSC SMVSC OPMC OMSC SFMC FPMVS EOMSC AWMVC FDAGF
Builds on: anchor graph multi-view clustering multi-view subspace clustering k-means
Closest papers by meaning
- Anchor Learning with Potential Cluster Constraints for Multi-view Clustering (AAAI 2025 · similarity 0.95)
- Towards Learnable Anchor for Deep Multi-View Clustering (AAAI 2025 · similarity 0.93)
- Max-Mahalanobis Anchors Guidance for Multi-View Clustering (AAAI 2025 · similarity 0.93)
- Learn from View Correlation: An Anchor Enhancement Strategy for Multi-view Clustering (CVPR 2024 · similarity 0.92)
- From Spectrum-free towards Baseline-view-free: Double-track Proximity Driven Multi-view Clustering (ICML 2025 · similarity 0.92)
- Robust Consensus Anchor Learning for Efficient Multi-view Subspace Clustering (ICML 2025 · similarity 0.92)
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.