Research map / Learning theory and algorithms / Clustering
From Dictionary to Tensor: A Scalable Multi-View Subspace Clustering Framework with Triple Information Enhancement
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 STONE, a scalable tensor-based multi-view subspace clustering framework that uses an enhanced anchor dictionary, an anchor hypergraph Laplacian regularizer, and a hyperbolic tangent rank surrogate for tensor rank.
Problem
Tensor-based multi-view subspace clustering (TMSC) methods have quadratic or cubic complexity, depend on observed data as a complete dictionary, ignore local geometry within views, and penalize all tensor singular values equally.
Approach
Replace self-representation with an enhanced anchor dictionary (anchor matrix plus unobserved samples, LatLRR-style) to get n×l anchor subspace representations, add an anchor hypergraph Laplacian for local geometry, and use a non-convex hyperbolic tangent rank that penalizes singular values adaptively instead of the tensor nuclear norm.
Key results
- On Scene15, STONE improves over the second-ranked method ETLMSC by 13% ACC and 13.7% ARI according to the authors.
- On ALOI-100, STONE runs in 66.57 seconds versus over 26000 seconds for TBGL.
- Ablation on Caltech101-all shows full STONE reaches ACC 0.994 and NMI 0.984, and removing HTR (STONE-v1) lowers ACC on all eight datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use STONE instead of the strongest listed baseline | Multi-view clustering · Scene15 | ACC: 0.847±0.030 -> 0.977±0.000 | Helped |
| use STONE instead of the strongest listed baseline | Multi-view clustering · NGs | ACC: 0.982±0.000 -> 1.000±0.000 | Helped |
| use STONE instead of the strongest listed baseline | Multi-view clustering · BBCSport | ACC: 0.961±0.000 -> 1.000±0.000 | Helped |
| use STONE instead of the strongest listed baseline | Multi-view clustering · HW | ACC: 0.998±0.000 -> 1.000±0.000 | Helped |
| use STONE instead of the strongest listed baseline | Multi-view clustering · MSRCV1 | ACC: 1.000±0.000 -> 1.000±0.000 | No clear effect |
| use STONE instead of the strongest listed baseline | Multi-view clustering · ALOI-100 | ACC: 0.767±0.000 -> 0.814±0.000 | Helped |
| use STONE instead of the strongest listed baseline | Multi-view clustering · Caltech101-all | ACC: 0.807±0.000 -> 0.650±0.000 | Hurt |
| use STONE instead of the strongest listed baseline | Multi-view clustering · CIFAR10 | ACC: 0.993±0.000 -> 0.994±0.000 | Helped |
| add HTR (ablation) | STONE · NGs | ACC: 0.379±0.000 -> 1.000±0.000 | Helped |
| add HTR (ablation) | STONE · Scene15 | ACC: 0.629±0.000 -> 0.977±0.000 | Helped |
| add LEAD loss (ablation) | STONE · NGs | ACC: 0.438 -> 1.000 | Helped |
| add LRE loss (ablation) | STONE · NGs | ACC: 0.208 -> 1.000 | Helped |
Surprising findings
- Shallow multi-view models such as STONE were reported to outperform deep methods MFLVC and GCFAgg in most scenarios.
- The authors report near-perfect clustering (ACC 1.000) on several datasets such as NGs, BBCSport, HW, and MSRCV1, which is unusually high relative to baselines.
Limitations
- The NeurIPS checklist marks limitations as NA and the paper does not discuss limitations.
- Balancing parameters alpha, beta, gamma and delta require grid search tuning, with performance fluctuating across values.
- Several reported results are exactly 1.000 with zero variance, raising concerns about evaluation.
Experimental setup
Datasets: NGs BBCSport HW Scene15 MSRCV1 Caltech101-all ALOI-100 CIFAR10
Benchmarks: NGs BBCSport HW Scene15 MSRCV1 Caltech101-all ALOI-100 CIFAR10
Metrics: ACC NMI PUR F-score ARI
Baselines: SMVSC SFMC GMC MSC2D MVCtopl MVSCTM ETLMSC TBGL MFLVC GCFAgg SC-best
Builds on: LRR LatLRR tensor nuclear norm anchor graph clustering ADMM
Closest papers by meaning
- S2MVTC: a Simple yet Efficient Scalable Multi-View Tensor Clustering (CVPR 2024 · similarity 0.91)
- EDISON: Enhanced Dictionary-Induced Tensorized Incomplete Multi-View Clustering with Gaussian Error Rank Minimization (ICML 2024 · similarity 0.90)
- Large-scale Multi-view Tensor Clustering with Implicit Linear Kernels (CVPR 2025 · similarity 0.90)
- Performance Gaps in Multi-view Clustering under the Nested Matrix-Tensor Model (ICLR 2024 · similarity 0.90)
- SLR-MVTC: Smooth Low-Rank Multi-View Tensor Clustering (AAAI 2025 · similarity 0.89)
- Contrastive Multi-view Subspace Clustering via Tensor Transformers Autoencoder (AAAI 2025 · similarity 0.88)
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.