Research map / Learning theory and algorithms / Clustering
Robust Consensus Anchor Learning for Efficient Multi-view Subspace 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 RCSC, an anchor-based multi-view subspace clustering method that jointly learns robust consensus anchors, anchor graph, and cluster indicators under Frobenius-norm, non-negative and affine constraints, with a block-diagonal theoretical analysis.
Problem
Multi-view subspace clustering on large-scale data is computationally expensive for graph methods; existing anchor-based methods do not guarantee robustness of learned consensus anchors nor jointly learn anchors and the shared cluster structure.
Approach
Learn a shared set of consensus anchors with view-specific orthogonal projections, impose Frobenius-norm, non-negative and affine constraints on the anchor-graph coefficients, and jointly optimize anchors, graph, and a discrete cluster indicator via orthogonal nonnegative factorization, with a theorem giving block-diagonal structure under independent subspaces.
Key results
- RCSC achieves the best ACC on all eight datasets in Table 1, e.g., 85.60% on Mfeat and 28.00% on Caltech101-all versus 27.50% for FPMVS.
- On AWA, RCSC's NMI is about 2.32% higher than MSGL, per the paper.
- Running time is comparable to the efficient MSGL baseline on most datasets, and the method is claimed to be nearly linear in n.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RCSC instead of OMSC | Multi-view clustering · ORL | ACC (%): 63.80±0.00 -> 65.20±0.00 | Helped |
| Use RCSC instead of OMSC | Multi-view clustering · Mfeat | ACC (%): 84.00±0.05 -> 85.60±0.00 | Helped |
| Use RCSC instead of FPMVS | Multi-view clustering · Caltech101-20 | ACC (%): 66.15±0.10 -> 67.40±0.00 | Helped |
| Use RCSC instead of FPMVS | Multi-view clustering · Caltech101-all | ACC (%): 27.50±0.05 -> 28.00±0.50 | Helped |
| Use RCSC instead of OMSC | Multi-view clustering · SUNRGBD | ACC (%): 25.20±0.00 -> 27.00±0.00 | Helped |
| Use RCSC instead of OMSC | Multi-view clustering · NUSWIDEOBJ | ACC (%): 21.00±0.05 -> 23.50±0.00 | Helped |
| Use RCSC instead of OMSC | Multi-view clustering · AWA | ACC (%): 9.00±0.10 -> 10.50±0.10 | Helped |
| Use RCSC instead of OMSC | Multi-view clustering · YoutubeFace | ACC (%): 26.50±0.00 -> 27.80±0.00 | Helped |
| Vary λ across the tested range (ablation) | RCSC · Eight multi-view datasets | Clustering performance (ACC, NMI, F1-score) | No clear effect |
| Vary β across the tested range (ablation) | RCSC · Eight multi-view datasets | Clustering performance (ACC, NMI, F1-score) | No clear effect |
| Vary the number of anchors from 2k to 7k (ablation) | RCSC · Eight multi-view datasets | Clustering performance (ACC, NMI, F1-score) | No clear effect |
Limitations
- Theorem 3 gives no sufficient condition for vertex bases, which the authors acknowledge.
- Improvements over the strongest baselines are often small, and some baselines report N/A due to out-of-memory failures on large datasets.
- Parameter sensitivity and anchor-number sweeps are shown only in figures, with tuning on the test benchmarks.
Experimental setup
Datasets: ORL Mfeat Caltech101-20 Caltech101-all SUNRGBD NUSWIDEOBJ AWA YoutubeFace
Benchmarks: ORL Mfeat Caltech101-20 Caltech101-all SUNRGBD NUSWIDEOBJ AWA YoutubeFace
Metrics: ACC NMI F1-score running time
Baselines: AMGL SFMC BMVC LMVSC MSGL FPMVS EOMSC-CA OMSC
Builds on: anchor graph multi-view clustering LMVSC MSGL
Closest papers by meaning
- Learning Cluster-Wise Anchors for Multi-View Clustering (AAAI 2024 · similarity 0.92)
- Scalable One-Pass Incomplete Multi-View Clustering by Aligning Anchors (AAAI 2025 · similarity 0.91)
- Max-Mahalanobis Anchors Guidance for Multi-View Clustering (AAAI 2025 · similarity 0.90)
- Anchor Learning with Potential Cluster Constraints for Multi-view Clustering (AAAI 2025 · similarity 0.90)
- Learn from View Correlation: An Anchor Enhancement Strategy for Multi-view Clustering (CVPR 2024 · similarity 0.89)
- Towards Learnable Anchor for Deep Multi-View Clustering (AAAI 2025 · similarity 0.89)
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.