Research map / Learning theory and algorithms / Clustering
From Spectrum-free towards Baseline-view-free: Double-track Proximity Driven 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 DTP-SF-BVF, an anchor-based multi-view clustering method that adds anchor-anchor self-expression structure, learnable per-view anchor permutations (no baseline view), and directly learns discrete cluster indicators without spectral embedding.
Problem
Anchor-based multi-view clustering ignores anchor-anchor relations, suffers anchor misalignment across views due to unknown cluster orders (and needs a baseline view to fix it), and produces unstable results because spectral embedding and partitioning introduce variance.
Approach
Jointly optimize anchor-anchor self-expression (used to guide the anchor-sample graph), learnable per-view permutation matrices that align anchors to a shared binary consensus cluster indicator, and direct binary cluster-indicator learning in place of spectrum construction, all in one alternating-optimization framework.
Key results
- On DERMATO, DTP-SF-BVF reaches 85.47% ACC, 89.97% NMI and 87.92% Fscore, the best among compared methods, and reports zero variance across runs.
- On CALTE7 the method achieves 80.66% ACC versus 71.98% for the best other method FMCNOF, a sizeable margin.
- Ablations show spectrum-free learning cuts runtime substantially (e.g., 3850s to 1194s on FasMNI4V) while the double-track proximity and baseline-view-free components each improve ACC over their ablated versions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add anchor-anchor proximity to anchor-sample proximity (ablation) | DTP-SF-BVF · DERMATO, CALTE7, Cora, REU7200, Reuters, CIF10Tra4, FasMNI4V | ACC (%): 71.51 -> 85.47; 49.05 -> 80.66; 30.35 -> 52.44; 16.75 -> 26.22; 47.05 -> 54.26; 26.69 -> 26.83; 52.15 -> 57.36 | Helped |
| use spectrum-free binary indicators instead of spectral clustering (ablation) | DTP-SF-BVF · DERMATO, CALTE7, Cora, REU7200, Reuters, CIF10Tra4, FasMNI4V | ACC (%): 81.33(±1.82) -> 85.47(±0.00); 74.74(±1.73) -> 80.66(±0.00); 48.27(±1.07) -> 52.44(±0.00); 23.46(±0.79) -> 26.22(±0.00); | Mixed |
| add baseline-view-free alignment (ablation) | DTP-SF-BVF · DERMATO, CALTE7, Cora, REU7200, Reuters, CIF10Tra4, FasMNI4V | ACC (%): 80.73 -> 85.47; 76.59 -> 80.66; 31.65 -> 52.44; 16.67 -> 26.22; 45.29 -> 54.26; 25.91 -> 26.83; 53.68 -> 57.36 | Helped |
| avoid selecting a baseline view (ablation) | DTP-SF-BVF · DERMATO, CALTE7, Cora, REU7200, Reuters, CIF10Tra4, FasMNI4V | ACC (%): 72.37 -> 85.47; 62.43 -> 80.66; 45.78 -> 52.44; 22.87 -> 26.22; 44.36 -> 54.26; 22.98 -> 26.83; 51.22 -> 57.36 | Helped |
Surprising findings
- On DERMATO, AEVC achieves higher ACC (91.82%) than the proposed method (85.47%), and on Cora/ REU7200 some baselines are close, so gains are not uniform across datasets.
Limitations
- Depends on two hyper-parameters (lambda, beta) that require tuning.
- Uses a square weighting scheme with linear constraints for view weights; other view-weighting schemes are unexplored.
- Permutation search for T_p relies on traversal over one-hot vectors, with factorial cost noted in complexity analysis.
Experimental setup
Datasets: DERMATO CALTE7 Cora REU7200 Reuters CIF10Tra4 FasMNI4V
Benchmarks: DERMATO CALTE7 Cora REU7200 Reuters CIF10Tra4 FasMNI4V
Metrics: ACC NMI Fscore running time
Baselines: FMR PMSC AMGL MSCIAS MVSC MLRSSC MPAC MCLES FMCNOF PFSC SFMC MSGL
Builds on: anchor-based multi-view clustering subspace self-expression FMVACC
Closest papers by meaning
- Learning Cluster-Wise Anchors for Multi-View Clustering (AAAI 2024 · similarity 0.92)
- Towards Learnable Anchor for Deep Multi-View Clustering (AAAI 2025 · similarity 0.91)
- Learn from View Correlation: An Anchor Enhancement Strategy for Multi-view Clustering (CVPR 2024 · similarity 0.91)
- Anchor Learning with Potential Cluster Constraints for Multi-view Clustering (AAAI 2025 · similarity 0.91)
- Max-Mahalanobis Anchors Guidance for Multi-View Clustering (AAAI 2025 · similarity 0.90)
- DVSAI: Diverse View-Shared Anchors Based Incomplete Multi-View Clustering (AAAI 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.