Research map / Learning theory and algorithms / Clustering
Evaluate then Cooperate: Shapley-based View Cooperation Enhancement for 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 SCE-MVC, which treats views as players in an unsupervised cooperative game, computes Shapley values with an optimal-transport-based utility, and uses them to modulate per-view training speeds to balance view contributions in deep multi-view clustering.
Problem
Deep multi-view clustering methods lack a way to measure how much each view contributes during fusion, and dominant views can suppress others, leading to suboptimal inter-view cooperation; unsupervised settings make contribution evaluation hard.
Approach
Define an label-free utility for view subsets as 1/(1+Wasserstein distance) between the subset's fused cluster distribution and a self-distilled target distribution, compute exact Shapley values over views, and use normalized contributions to set per-view gradient scaling factors k_v = exp(tau(phi_min/phi_v - 1)) so that dominant views learn more slowly.
Key results
- On Caltech101-7, DMJC+SCE raises ACC from 0.469 to 0.583 (NMI 0.411 to 0.513, ARI 0.309 to 0.457).
- On STL10, DMJC+SCE reaches ACC 0.938 vs 0.893 for DMJC without SCE, and on UCI-digit ACC 0.927 vs 0.871.
- Shapley-based contribution of the suppressed second view on Caltech101-7 rises from 0.032 to 0.433 after applying SCE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SCE to DMJC (ablation) | DMJC · Caltech101-7 | ACC: 0.469 -> 0.583 | Helped |
| Add SCE to DMJC (ablation) | DMJC · Caltech101-7 | ARI: 0.309 -> 0.457 | Helped |
| Add SCE to DMJC (ablation) | DMJC · CUB | ACC: 0.758 -> 0.797 | Helped |
| Add SCE to DMJC (ablation) | DMJC · UCI-digit | ACC: 0.871 -> 0.927 | Helped |
| Add SCE to DMJC (ablation) | DMJC · UCI-digit | ARI: 0.767 -> 0.847 | Helped |
| Add SCE to DMJC (ablation) | DMJC · HandWritten | ACC: 0.893 -> 0.938 | Helped |
| Add SCE to DMJC (ablation) | DMJC · STL10 | ACC: 0.305 -> 0.330 | Helped |
| Add SCE to DMJC (ablation) | DMJC · Reuters | ACC: 0.485 -> 0.533 | Helped |
| Add SCE to DMJC (ablation) | DMJC · Reuters | ARI: 0.241 -> 0.237 | Hurt |
| Add SCE to InfoNCE+Kmeans (ablation) | InfoNCE+Kmeans · CUB | ACC: 0.715 -> 0.717 | No clear effect |
| Add SCE to ProIMP (ablation) | ProIMP · CUB | ACC: 0.825 -> 0.832 | Helped |
| Add SCE to ProIMP (ablation) | ProIMP · Caltech101-7 | ACC: 0.382 -> 0.382 | No clear effect |
Surprising findings
- Some view contributions computed by the evaluation module can be negative on extreme datasets, contradicting the non-negativity of Shapley values.
- For alignment-based methods (InfoNCE+Kmeans, ProImp), SCE yields only subtle gains because view contributions are already nearly equal.
Limitations
- Some view contributions can be negative on extreme datasets, which conflicts with Shapley non-negativity and raises the question of whether such views should be discarded.
- The cooperation module only narrows contribution gaps through gradient modulation and cannot amplify contributions of weak views.
- No error bars or variance across runs are reported, and gains on alignment-based methods are small.
Experimental setup
Datasets: CUB Caltech101-7 HandWritten UCI-digit STL10 Reuters WikipediaArticles SCENE15 NoisyMNIST
Benchmarks: Caltech101-7 CUB UCI-digit HandWritten STL10 Reuters WikipediaArticles SCENE15 NoisyMNIST
Metrics: ACC NMI ARI
Baselines: DEMVC CoMVC SiMVC SDSNE MFLVC SDMVC DSMVC APADC DMJC ProImp
Builds on: DEC Shapley value InfoNCE ProImp DMJC
Closest papers by meaning
- Mixture of Experts as Representation Learner for Deep Multi-View Clustering (AAAI 2025 · similarity 0.88)
- EASEMVC:Efficient Dual Selection Mechanism for Deep Multi-View Clustering (CVPR 2025 · similarity 0.87)
- Deep Fair Multi-View Clustering with Attention KAN (CVPR 2025 · similarity 0.86)
- Investigating and Mitigating the Side Effects of Noisy Views for Self-Supervised Clustering Algorithms in Practical Multi-View Scenarios (CVPR 2024 · similarity 0.86)
- Deep Variational Incomplete Multi-View Clustering: Exploring Shared Clustering Structures (AAAI 2024 · similarity 0.86)
- Medusa: A Multi-Scale High-order Contrastive Dual-Diffusion Approach for Multi-View Clustering (CVPR 2025 · similarity 0.86)
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.