Research map / Learning theory and algorithms / Clustering
Multi-View Clustering by Inter-cluster Connectivity Guided Reward
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 a non-parametric multi-view clustering method that infers the unknown cluster number k via a multi-armed bandit whose reward is the negative inter-cluster connectivity on a consensus mutual-nearest-neighbor graph.
Problem
Multi-view clustering usually requires a predefined number of clusters k, which is rarely known in real-world settings.
Approach
Treat each candidate k as a bandit arm, updated by epsilon-greedy with optimistic initialization, and reward it by the negative inter-cluster connectivity of the indicator matrix over a consensus MNN graph, which is argued to be minimized at the true k and to be stable over iterations.
Key results
- On BBC14, the method reports NMI 87.72, Purity 99.31, ARI 94.11, outperforming baselines by about 13.64%, 2.76%, and 4.60% in NMI, Purity and ARI.
- On Handwritten-numerals, NMI 94.38 and Purity 97.50 versus best baseline MLAN NMI 90.40 and Purity 95.60.
- On MSRC-v1 the method reports the highest NMI (79.76) and Purity (88.57) but a much lower ARI (74.51) than baselines such as CoMSC (91.77).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed inter-cluster-connectivity-guided method instead of the strongest reported baseline | Multi-view clustering · MSRC-v1 | NMI: 73.90 ± 0.00 -> 79.76 ± 0.00 | Helped |
| Use the proposed inter-cluster-connectivity-guided method instead of the strongest reported baseline | Multi-view clustering · MSRC-v1 | ARI: 91.77 ± 0.27 -> 74.51 ± 0.00 | Hurt |
| Use the proposed inter-cluster-connectivity-guided method instead of the strongest reported baseline | Multi-view clustering · HW | NMI: 90.40 ± 0.00 -> 94.38 ± 0.00 | Helped |
| Use the proposed inter-cluster-connectivity-guided method instead of the strongest reported baseline | Multi-view clustering · HW | ARI: 98.29 ± 0.00 -> 94.46 ± 0.00 | Hurt |
| Use the proposed inter-cluster-connectivity-guided method instead of the strongest reported baseline | Multi-view clustering · BBC14 | NMI: 74.08 ± 0.00 -> 87.72 ± 0.00 | Helped |
| Use the proposed inter-cluster-connectivity-guided method instead of the strongest reported baseline | Multi-view clustering · BBC14 | Purity: 96.55 ± 0.00 -> 99.31 ± 0.00 | Helped |
| Use the proposed inter-cluster-connectivity-guided method instead of the strongest reported baseline | Multi-view clustering · ORL | NMI: 89.89 ± 0.00 -> 90.39 ± 0.00 | Helped |
| Use the proposed inter-cluster-connectivity-guided method instead of the strongest reported baseline | Multi-view clustering · ORL | Purity: 92.50 ± 0.00 -> 82.25 ± 0.00 | Hurt |
| Use the proposed inter-cluster-connectivity-guided method instead of the strongest reported baseline | Multi-view clustering · ORL | ARI: 98.71 ± 0.00 -> 69.18 ± 0.00 | Hurt |
Surprising findings
- On MSRC-v1 and ORL, the method's ARI is substantially lower than several baselines (e.g., 74.51 vs 91.77 on MSRC-v1; 69.18 vs ~98 on ORL) despite higher NMI/Purity on some datasets.
- Clustering quality of the method across alpha, beta, gamma settings identifies the optimal k in only 37 of 50 settings.
Limitations
- Reward depends on the quality of the MVC indicator matrix and its hyperparameters; optimal k identified in only 37 of 50 settings.
- Reward analysis assumes stationarity; fluctuations remain in convergence curves.
- Reported ARI is substantially lower than some baselines on several datasets.
Experimental setup
Datasets: MSRC-v1 Handwritten-numerals bbc-seg14of4 ORL
Benchmarks: MSRC-v1 Handwritten-numerals BBC14 ORL
Metrics: NMI Purity ARI
Baselines: Co-train Co-reg MLAN AWP mPAC GMC SMVSC FPMVS COMVSC CSMSC PMSC CoMSC
Builds on: Liu et al. 2023 DQN-based cluster number inference spectral clustering COMIC
Closest papers by meaning
- A Non-parametric Graph Clustering Framework for Multi-View Data (AAAI 2024 · similarity 0.88)
- Medusa: A Multi-Scale High-order Contrastive Dual-Diffusion Approach for Multi-View Clustering (CVPR 2025 · similarity 0.88)
- Towards Learnable Anchor for Deep Multi-View Clustering (AAAI 2025 · similarity 0.88)
- Learning Cluster-Wise Anchors for Multi-View Clustering (AAAI 2024 · similarity 0.88)
- EASEMVC:Efficient Dual Selection Mechanism for Deep Multi-View Clustering (CVPR 2025 · similarity 0.88)
- Learn from View Correlation: An Anchor Enhancement Strategy for Multi-view Clustering (CVPR 2024 · 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.