Research map / Learning theory and algorithms / Clustering
Learn from View Correlation: An Anchor Enhancement Strategy 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 AEVC, a plug-and-play anchor enhancement strategy that aligns initial anchor graphs across views, builds a view graph, and reconstructs anchors from neighboring views to improve anchor-based multi-view clustering.
Problem
Anchor-based multi-view clustering performance depends heavily on anchor quality, but existing anchor generation uses only single-view information and iterative anchor optimization remains sensitive to initialization.
Approach
Build a view graph from aligned anchor graphs and enhance each view's anchors using anchor-sample relations from its neighboring (similar) views, so anchors across similar views have tighter distributions; then construct a view-weighted anchor graph guided by a concatenated-view graph.
Key results
- AEVC outperforms eight anchor-based MVC baselines on all seven datasets; ACC gains over the second-best method range from 1.3% to 15.8%.
- On Reuters, the anchor enhancement module improves FMCNOF, LMVSC, UDBGL and FMVACC by 18.4%, 6.1%, 27.8% and 21.7% respectively.
- Ablation on Dermatology: baseline 90.87 ACC, +alignment 91.82, +alignment+enhancement 93.20.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply AEVC instead of the strongest baseline | anchor-based multi-view clustering · Dermatology | ACC: 87.15±0.00 -> 93.20±4.50 | Helped |
| apply AEVC instead of the strongest baseline | anchor-based multi-view clustering · ForestTypes | ACC: 79.29±0.23 -> 82.53±0.09 | Helped |
| apply AEVC instead of the strongest baseline | anchor-based multi-view clustering · BDGP | ACC: 59.51±4.00 -> 60.65±0.62 | Helped |
| apply AEVC instead of the strongest baseline | anchor-based multi-view clustering · Reuters | ACC: 27.22±0.24 -> 29.91±0.02 | Helped |
| apply AEVC instead of the strongest baseline | anchor-based multi-view clustering · Hdigit | ACC: 86.87±0.08 -> 88.01±2.97 | Helped |
| apply AEVC instead of the strongest baseline | anchor-based multi-view clustering · VGGFace2 | ACC: 7.21±0.00 -> 8.35±0.07 | Helped |
| apply AEVC instead of the strongest baseline | anchor-based multi-view clustering · CIFAR100 | ACC: 9.81±0.00 -> 10.74±0.08 | Helped |
| add anchor alignment (ablation) | AEVC · Dermatology, ForestTypes, BDGP | ACC: 90.87±3.60 -> 91.82±3.78; 70.76±0.05 -> 75.51±0.17; 54.38±0.10 -> 57.97±0.09 | Helped |
| add anchor enhancement after alignment (ablation) | AEVC · Dermatology, ForestTypes, BDGP | ACC: 91.82±3.78 -> 93.20±4.50; 75.51±0.17 -> 82.53±0.09; 57.97±0.09 -> 60.65±0.62 | Helped |
| add view weights to anchor graph construction (ablation) | AEVC · Dermatology, ForestTypes, BDGP | ACC: 91.60±6.26 -> 93.20±4.50; 81.45±0.00 -> 82.53±0.09; 60.50±0.46 -> 60.65±0.62 | Helped |
| add the regularization term (ablation) | AEVC · Dermatology, ForestTypes, BDGP | ACC: 89.98±3.40 -> 93.20±4.50; 79.75±0.10 -> 82.53±0.09; 60.64±0.62 -> 60.65±0.62 | Helped |
| integrate the anchor enhancement module (ablation) | FMCNOF · Dermatology, ForestTypes, BDGP, Reuters, Hdigit, VGGFace2, CIFAR100 | ACC: 62.01±0.00 -> 73.18±0.00; 45.51±0.00 -> 69.60±0.00; 31.08±0.00 -> 37.64±0.00; 19.25±0.00 -> 22.79±0.00; 32.84±0.00 -> 34 | Mixed |
Surprising findings
- Adding the enhancement module does not help every method; UDBGL-AE drops on Dermatology (84.08 to 81.01 ACC).
Limitations
- Enhancement rate gamma is sensitive; extreme values disrupt balance between new and original anchors.
- Hyperparameters (m, gamma, lambda) are chosen by grid search over predefined lists.
Experimental setup
Datasets: Dermatology ForestTypes BDGP Reuters Hdigit VGGFace2 CIFAR100
Benchmarks: Dermatology ForestTypes BDGP Reuters Hdigit VGGFace2 CIFAR100
Metrics: ACC NMI Purity F-score
Baselines: LMVSC SMVSC FMCNOF FPMVS-CAG UDBGL FastMICE FDAGF FMVACC
Builds on: anchor graph multi-view clustering LMVSC UDBGL FMVACC
Closest papers by meaning
- Max-Mahalanobis Anchors Guidance for Multi-View Clustering (AAAI 2025 · similarity 0.94)
- Learning Cluster-Wise Anchors for Multi-View Clustering (AAAI 2024 · similarity 0.92)
- Anchor Learning with Potential Cluster Constraints for Multi-view Clustering (AAAI 2025 · similarity 0.92)
- Alleviate Anchor-Shift: Explore Blind Spots with Cross-View Reconstruction for Incomplete Multi-View Clustering (NeurIPS 2024 · similarity 0.92)
- Fast Incomplete Multi-view Clustering by Flexible Anchor Learning (ICML 2025 · similarity 0.91)
- From Spectrum-free towards Baseline-view-free: Double-track Proximity Driven Multi-view Clustering (ICML 2025 · similarity 0.91)
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.