Research map / Graph learning / Graph algorithms and clustering
Towards Learnable Anchor for Deep 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 DMAC, a deep multi-view clustering model that makes anchors learnable via Gaussian positive-incentive perturbations, anchor graph convolution with cross-view mutual information maximization, achieving linear-time clustering.
Problem
Deep multi-view clustering with graph learning costs O(n^2) time; anchor-based alternatives use fixed, manually selected anchors that may not reflect the true cluster distribution, corrupting clustering.
Approach
Inject learnable perturbations (reparameterized Gaussian noise, treated as positive-incentive noise) into initial k-means anchors, optimize them with an anchor learning entropy loss, and derive cluster-oriented anchors via anchor graph convolution and cross-view mutual information maximization, then build the full sample graph from the learned anchors for a structure-preserving embedding.
Key results
- DMAC achieves average ACC 55.98 and NMI 57.40 across six datasets, versus 45.17/50.52 for the best competitor (GCFAggMVC ACC, SURER NMI).
- On Yale, DMAC reaches ACC 78.18 vs 75.18 for DealMVC; on ALOI, ACC 60.35 vs 58.51 for LMVSC.
- Ablation: removing the perturbation-driven module (wo/PD) drops ALOI ACC from 60.35 to 44.45 and PIE ACC from 43.24 to 34.12.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add perturbation generation network and anchor learning loss (ablation) | DMAC · Yale | ACC: 66.02 -> 78.18 | Helped |
| add perturbation generation network and anchor learning loss (ablation) | DMAC · PIE | ACC: 34.12 -> 43.24 | Helped |
| add perturbation generation network and anchor learning loss (ablation) | DMAC · BBC | ACC: 87.15 -> 88.61 | Helped |
| add perturbation generation network and anchor learning loss (ablation) | DMAC · NUS | ACC: 24.53 -> 29.29 | Helped |
| add perturbation generation network and anchor learning loss (ablation) | DMAC · CCV | ACC: 32.28 -> 36.18 | Helped |
| add perturbation generation network and anchor learning loss (ablation) | DMAC · ALOI | ACC: 44.45 -> 60.35 | Helped |
| add consistency maximization loss (ablation) | DMAC · Yale | ACC: 72.73 -> 78.18 | Helped |
| add consistency maximization loss (ablation) | DMAC · PIE | ACC: 33.82 -> 43.24 | Helped |
| add consistency maximization loss (ablation) | DMAC · BBC | ACC: 88.47 -> 88.61 | Helped |
| add consistency maximization loss (ablation) | DMAC · NUS | ACC: 27.67 -> 29.29 | Helped |
| add consistency maximization loss (ablation) | DMAC · CCV | ACC: 34.26 -> 36.18 | Helped |
| add consistency maximization loss (ablation) | DMAC · ALOI | ACC: 54.79 -> 60.35 | Helped |
Surprising findings
- On NUS, DMAC's ACC (29.29) is below DFP-GNN (29.42), and on ALOI its NMI is below LMVSC, so it is not best on every metric.
Limitations
- Anchors are initialized by k-means on the fused embedding, so the initial anchor quality still depends on a manual discretization step.
- Performance on NUS and some NMI metrics is not uniformly best; gains on several datasets are small.
Experimental setup
Datasets: Yale PIE BBC NUS CCV ALOI
Benchmarks: Yale PIE BBC NUS CCV ALOI
Metrics: ACC NMI runtime
Baselines: GMC MSGL LMVSC UDBGL CMGEC DealMVC GCFAggMVC DFP-GNN SURER
Builds on: positive-incentive noise anchor graph learning graph convolution reparameterization trick
Closest papers by meaning
- Anchor Learning with Potential Cluster Constraints for Multi-view Clustering (AAAI 2025 · similarity 0.93)
- Learning Cluster-Wise Anchors for Multi-View Clustering (AAAI 2024 · similarity 0.93)
- Max-Mahalanobis Anchors Guidance for Multi-View Clustering (AAAI 2025 · similarity 0.92)
- From Spectrum-free towards Baseline-view-free: Double-track Proximity Driven Multi-view Clustering (ICML 2025 · similarity 0.91)
- Learn from View Correlation: An Anchor Enhancement Strategy for Multi-view Clustering (CVPR 2024 · similarity 0.91)
- Alleviate Anchor-Shift: Explore Blind Spots with Cross-View Reconstruction for Incomplete Multi-View Clustering (NeurIPS 2024 · 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.