Research map / Learning theory and algorithms / Clustering
DVSAI: Diverse View-Shared Anchors Based Incomplete 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 DVSAI, an incomplete multi-view clustering method that generates view-shared anchors with multiple dimensions and sizes across several potential spaces, with a linear-cost graph integration scheme and an iterative four-variable solver.
Problem
Incomplete multi-view clustering, where samples are missing in some views; existing anchor-based view-shared methods use anchors with only one dimension and one size, limiting diversity and clustering quality, and many IMVC methods do not scale to large datasets.
Approach
Associates each partial view with several potential spaces (projection matrices P_v,t) so view-shared anchors get space-specific dimension and size; integrates the multi-scale unified anchor graphs via a linear-cost scheme (SVD of a low-rank matrix L) to get the spectral embedding.
Key results
- On Caltech101-20 at missing rate 0.1, DVSAI reaches 71.14 ACC, 68.63 NMI, 79.91 PUR, exceeding all compared IMVC methods.
- On YoutubeFace Sel (>100k samples), most baselines run out of memory while DVSAI runs and gives the best results on three metrics.
- Ablation over potential spaces shows T=5 gives the best results in most settings; the integration scheme is much faster than COMP (e.g., 151.21s vs 4.32s on NUSWIDEOBJ at missing rate 0.1).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use diverse view-shared anchors with multi-dimension and multi-size instead of single-dimension single-size anchors (ablation) | DVSAI · Caltech101-7, missing rate 0.1 | ACC: 51.22 -> 62.73 | Helped |
| Use diverse view-shared anchors with multi-dimension and multi-size instead of single-dimension multi-size anchors (ablation) | DVSAI · Caltech101-7, missing rate 0.1 | ACC: 53.12 -> 62.73 | Helped |
| Use diverse view-shared anchors with multi-dimension and multi-size instead of multi-dimension single-size anchors (ablation) | DVSAI · Caltech101-7, missing rate 0.1 | ACC: 53.91 -> 62.73 | Helped |
| Increase the number of potential spaces from 1 to 5 (ablation) | DVSAI · Caltech101-7, missing rate 0.1 | ACC: 51.22 -> 62.73 | Helped |
| Increase the number of potential spaces from 2 to 5 (ablation) | DVSAI · Caltech101-7, missing rate 0.1 | ACC: 52.31 -> 62.73 | Helped |
| Increase the number of potential spaces from 3 to 5 (ablation) | DVSAI · Caltech101-7, missing rate 0.1 | ACC: 61.46 -> 62.73 | Helped |
| Increase the number of potential spaces from 4 to 5 (ablation) | DVSAI · Caltech101-7, missing rate 0.1 | ACC: 62.15 -> 62.73 | Helped |
| Use DVSAI rather than the strongest listed baseline | Caltech101-7 · Caltech101-7, missing rate 0.1 | ACC: 62.65 -> 62.73 | Helped |
| Use DVSAI rather than the strongest listed baseline | Caltech101-20 · Caltech101-20, missing rate 0.1 | ACC: 52.44 -> 71.14 | Helped |
| Use DVSAI rather than the strongest listed baseline | CCV · CCV, missing rate 0.1 | ACC: 21.73 -> 23.67 | Helped |
| Use DVSAI rather than the strongest listed baseline | SUNRGBD · SUNRGBD, missing rate 0.1 | ACC: 20.74 -> 23.98 | Helped |
| Use DVSAI rather than the strongest listed baseline | NUSWIDEOBJ · NUSWIDEOBJ, missing rate 0.1 | ACC: 12.35 -> 18.85 | Helped |
Surprising findings
- BGIMVSC, PIMVC and HCLS are slightly better than DVSAI in some metric/dataset settings.
- Using more potential spaces does not always help; some results are sub-optimal at T=5, attributed to redundancy among spaces.
Limitations
- Number of potential spaces T must be set manually; automatic determination is left for future work.
- Some baselines (BGIMVSC, PIMVC, HCLS) outperform DVSAI in certain settings.
- Many baselines run out of memory on large datasets, so comparisons are incomplete there.
Experimental setup
Datasets: Caltech101-7 Caltech101-20 CCV SUNRGBD NUSWIDEOBJ YoutubeFace Sel
Benchmarks: Caltech101-7 Caltech101-20 CCV SUNRGBD NUSWIDEOBJ YoutubeFace Sel
Metrics: ACC NMI PUR execution time
Baselines: BSV MIC MKKM-IK UEAF IK-MKC FLSD IMVC-CBG LSIMVC BGIMVSC PIMVC HCLS
Builds on: anchor graph clustering view-shared anchor framework COMP graph integration
Closest papers by meaning
- Towards Resource-friendly, Extensible and Stable Incomplete Multi-view Clustering (ICML 2024 · similarity 0.92)
- Fast Incomplete Multi-view Clustering by Flexible Anchor Learning (ICML 2025 · similarity 0.92)
- Scalable One-Pass Incomplete Multi-View Clustering by Aligning Anchors (AAAI 2025 · similarity 0.91)
- Fast Incomplete Multi-view Clustering with Adaptive Similarity Completion and Reconstruction (AAAI 2025 · similarity 0.91)
- From Spectrum-free towards Baseline-view-free: Double-track Proximity Driven Multi-view Clustering (ICML 2025 · similarity 0.90)
- Learn from View Correlation: An Anchor Enhancement Strategy for Multi-view Clustering (CVPR 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.