Research map / Graph learning / Graph algorithms and clustering
Anchor Learning with Potential Cluster Constraints 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 ALPC, an anchor-based multi-view clustering method that learns anchors dynamically under a shared potential cluster constraint so anchors are spread across data clusters, jointly optimized with anchor graph construction.
Problem
Dynamically learned anchors in anchor-based multi-view clustering are unevenly distributed, with some lying outside data clusters or clusters lacking anchors, reducing anchor quality and clustering performance.
Approach
Introduces a shared latent cluster-structure constraint that aligns anchor cluster centers (via a basis matrix and anchor-cluster indicator P) with the cluster indicator R of raw samples, so anchors are generated from data clusters, combined with orthogonality constraints in a unified alternating optimization.
Key results
- ALPC reports ACC 0.8819 on MSRC and 0.9512 on BBCSport, exceeding the second-best method by 10.09% and 12.58% respectively.
- On Wiki, ALPC achieves ACC 0.6040 and NMI 0.5514, versus EOMSC's 0.5415 ACC and 0.5290 NMI, though FDAGF has higher Purity (0.6811 vs 0.6399).
- Ablation (Ours-a without shared cluster constraint) shows ALPC better on all six datasets, e.g., MNIST ACC 1.0000 vs 0.9108.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add shared potential-cluster constraints and unified anchor-graph learning | ALPC · MSRC | ACC: 0.7810 -> 0.8819 | Helped |
| Add shared potential-cluster constraints and unified anchor-graph learning | ALPC · BBCSport | ACC: 0.8254 -> 0.9512 | Helped |
| Add shared potential-cluster constraints and unified anchor-graph learning | ALPC · Wiki | ACC: 0.5415 -> 0.6040 | Helped |
| Add shared potential-cluster constraints and unified anchor-graph learning | ALPC · Caltech101-all | ACC: 0.3015 -> 0.3284 | Helped |
| Add shared potential-cluster constraints and unified anchor-graph learning | ALPC · MNIST | ACC: 0.9905 -> 1.0000 | Helped |
| Add shared potential-cluster constraints and unified anchor-graph learning | ALPC · YouTubeFace_sel | ACC: 0.2414 -> 0.3046 | Helped |
| Add shared latent clustering structure (ablation) | ALPC · MSRC | ACC: 0.7507 -> 0.8819 | Helped |
| Add shared latent clustering structure (ablation) | ALPC · BBCSport | ACC: 0.8956 -> 0.9512 | Helped |
| Add shared latent clustering structure (ablation) | ALPC · Wiki | ACC: 0.5201 -> 0.6040 | Helped |
| Add shared latent clustering structure (ablation) | ALPC · Caltech101-all | ACC: 0.1007 -> 0.3284 | Helped |
| Add shared latent clustering structure (ablation) | ALPC · MNIST | ACC: 0.9108 -> 1.0000 | Helped |
| Add shared latent clustering structure (ablation) | ALPC · YouTubeFace_sel | ACC: 0.1530 -> 0.3046 | Helped |
Surprising findings
- On Wiki, the proposed method has lower Purity (0.6399) than FDAGF (0.6811), contrary to its overall claims.
- On Caltech101-all, the best NMI and purity reflect large gains, yet ACC is only 0.3284, indicating the task remains hard.
Limitations
- Hyperparameters lambda1, lambda2 and anchor count m need grid search and are sensitive, especially lambda2.
- Convergence argument is only partial since the joint problem is non-convex and the monotone-decrease claim is loosely justified.
- Reported gains on some datasets (Wiki Purity, Caltech F-score) are mixed against baselines.
Experimental setup
Datasets: MSRC BBCSport Wiki Caltech101-all MNIST YouTubeFace_sel
Benchmarks: MSRC BBCSport Wiki Caltech101-all MNIST YouTubeFace_sel
Metrics: ACC NMI Purity F-score
Baselines: LMVSC SMVSC OPMC OMSC SFMC FPMVS EOMSC AWMVC FDAGF
Builds on: anchor graph multi-view clustering multi-view subspace clustering k-means anchor selection
Closest papers by meaning
- Learning Cluster-Wise Anchors for Multi-View Clustering (AAAI 2024 · similarity 0.95)
- Max-Mahalanobis Anchors Guidance for Multi-View Clustering (AAAI 2025 · similarity 0.94)
- Towards Learnable Anchor for Deep 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 0.91)
- Robust Consensus Anchor Learning for Efficient Multi-view Subspace Clustering (ICML 2025 · 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.