Research map / Graph learning / Graph algorithms and clustering
Tensorized Label Learning on Anchor Graph
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 TLL-AG, a multi-view clustering model that factorizes per-view anchor graphs with orthogonal non-negative matrix factorization to get soft labels directly, and regularizes the stacked label tensor with tensor Schatten p-norm, optimized via ALM.
Problem
Graph-based multimedia/multi-view clustering needs costly post-processing to obtain labels from anchor graphs, and existing anchor-graph methods align cross-view labels pixel-by-pixel (F-norm), ignoring spatial/cross-view structure.
Approach
Decompose each view's anchor graph via G-orthogonal non-negative matrix factorization so the indicator matrix directly yields cluster labels, and impose a tensor Schatten p-norm on the stacked label tensor to capture complementary cross-view information.
Key results
- On MSRC, the method reaches ACC 0.986 and NMI 0.968, versus 0.981 ACC for the second-best MSC-BG.
- On Mnist4, ACC 0.981 and NMI 0.935 vs. 0.938 ACC for MSC-BG... ETLMSC reports 0.934 ACC, so the method leads all listed baselines.
- On NoisyMNIST (50,000 samples), ACC 0.605 vs. 0.388 for LMVSC; SMSC, ETLMSC and MSC-BG run out of memory on large datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TLL-AG instead of MSC-BG | Multi-view clustering · MSRC | ACC: 0.981 -> 0.986 | Helped |
| Use TLL-AG instead of MSC-BG | Multi-view clustering · MSRC | NMI: 0.960 -> 0.968 | Helped |
| Use TLL-AG instead of ETLMSC | Multi-view clustering · HandWritten4 | ACC: 0.938 -> 0.982 | Helped |
| Use TLL-AG instead of MSC-BG | Multi-view clustering · HandWritten4 | NMI: 0.922 -> 0.962 | Helped |
| Use TLL-AG instead of MSC-BG | Multi-view clustering · Mnist4 | ACC: 0.938 -> 0.981 | Helped |
| Use TLL-AG instead of MSC-BG | Multi-view clustering · Mnist4 | NMI: 0.861 -> 0.935 | Helped |
| Use TLL-AG instead of MSC-BG | Multi-view clustering · Reuters | ACC: 0.640 -> 0.686 | Helped |
| Use TLL-AG instead of MSC-BG | Multi-view clustering · Reuters | NMI: 0.484 -> 0.554 | Helped |
| Use TLL-AG instead of FPMVS-CAG | Multi-view clustering · NoisyMNIST | ACC: 0.554 -> 0.605 | Helped |
| Use TLL-AG instead of FPMVS-CAG | Multi-view clustering · NoisyMNIST | NMI: 0.513 -> 0.593 | Helped |
| Vary the anchor rate (ablation) | TLL-AG · HandWritten4 and Mnist4 | Clustering metrics | Mixed |
| Vary the anchor rate (ablation) | TLL-AG · HandWritten4 and Mnist4 | Running time | Hurt |
Surprising findings
- Several baselines (SMSC, ETLMSC, MSC-BG) run out of memory on the large-scale Reuters and NoisyMNIST datasets.
Limitations
- Performance is sensitive to the anchor rate and to the hyperparameters beta and p, which must be tuned.
- Assumes shared cluster label structure across views via tensor low-rank regularization; the paper notes FMCNOF's similar assumption may over-fit.
Experimental setup
Datasets: MSRC HandWritten4 Mnist4 Reuters NoisyMNIST
Benchmarks: MSRC HandWritten4 Mnist4 Reuters NoisyMNIST
Metrics: ACC NMI Purity running time
Baselines: FMCNOF LMVSC GMC SMSC FPMVS-CAG ETLMSC MSC-BG
Builds on: FMCNOF anchor graph tensor Schatten p-norm G-orthogonal NMF augmented Lagrange multiplier
Closest papers by meaning
- Tensorized Label Learning Based Fast Fuzzy Clustering (AAAI 2025 · similarity 0.89)
- Tensorized Label Learning via Balanced Tensor Regression (AAAI 2026 · similarity 0.88)
- Embedded Feature Selection on Graph-Based Multi-View Clustering (AAAI 2024 · similarity 0.86)
- Large-scale Multi-view Tensor Clustering with Implicit Linear Kernels (CVPR 2025 · similarity 0.85)
- S2MVTC: a Simple yet Efficient Scalable Multi-View Tensor Clustering (CVPR 2024 · similarity 0.84)
- From Dictionary to Tensor: A Scalable Multi-View Subspace Clustering Framework with Triple Information Enhancement (NeurIPS 2024 · similarity 0.84)
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.