Research map / Learning theory and algorithms / Clustering
Mixture of Experts as Representation Learner 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 DMVC-CE, a mixture-of-experts representation learner with top-K gating, an equilibrium loss, and a kernel-based distinctiveness enhancer for deep multi-view clustering, evaluated on five multi-view benchmarks.
Problem
Deep multi-view clustering methods either use separate per-view feature extractors that limit cross-view collaboration or a single shared module that fails to capture view-specific representations.
Approach
Uses a mixture-of-experts as a unified representation learner in which a gating network dynamically selects top-K experts per sample across views, regularized by an equilibrium loss for balanced expert usage and a pairwise Gaussian-kernel distinctiveness loss for expert diversity.
Key results
- DMVC-CE reaches 91.54% ACC on ALOI and 98.75% ACC on HW, outperforming the listed MVC baselines.
- On NUS-WIDE, DMVC-CE surpasses the GCL baseline by 14.20% ACC, 27.65% NMI, and 19.50% ARI as reported in the text.
- Ablation on ALOI and HW shows removing both the equilibrium loss and distinctiveness enhancer drops ACC to 80.82% and 87.00%, and replacing the MoE with multiple separate extractors gives 54.90% and 64.38%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DMVC-CE instead of the strongest reported baseline | MVC methods · ALOI | ACC: 83.43±2.41 -> 91.54±0.59 | Helped |
| Use DMVC-CE instead of the strongest reported baseline | MVC methods · Caltech101-all | ACC: 36.65±0.00 -> 42.54±1.99 | Helped |
| Use DMVC-CE instead of the strongest reported baseline | MVC methods · NUS-WIDE | ACC: 39.88±2.50 -> 54.08±1.03 | Helped |
| Use DMVC-CE instead of the strongest reported baseline | MVC methods · HW | ACC: 83.45±2.31 -> 98.75±1.01 | Helped |
| Use DMVC-CE instead of the strongest reported baseline | MVC methods · BDGP | ACC: 75.92±0.00 -> 80.82±0.30 | Helped |
| Replace separate view-specific extractors with a shared MoE learner (ablation) | DMVC-CE · ALOI | ACC: 54.90±2.42 -> 91.54±0.59 | Helped |
| Add equilibrium loss to balance expert selection (ablation) | DMVC-CE · ALOI | ACC: 87.67±1.71 -> 91.54±0.59 | Helped |
| Add distinctiveness enhancer to encourage expert diversity (ablation) | DMVC-CE · ALOI | ACC: 84.02±4.19 -> 91.54±0.59 | Helped |
| Add equilibrium loss and distinctiveness enhancer (ablation) | DMVC-CE · ALOI | ACC: 80.82±5.43 -> 91.54±0.59 | Helped |
| Add equilibrium loss to balance expert selection (ablation) | DMVC-CE · HW | ACC: 93.97±4.52 -> 98.75±1.01 | Helped |
| Add distinctiveness enhancer to encourage expert diversity (ablation) | DMVC-CE · HW | ACC: 93.37±5.10 -> 98.75±1.01 | Helped |
Limitations
- Two-stage framework: representation learning followed by k-means rather than end-to-end clustering.
- Hyperparameters lambda and gamma require tuning over a grid; performance depends on the chosen range.
- Results reported only on five benchmark datasets with small sample sizes.
Experimental setup
Datasets: ALOI Caltech101-all NUS-WIDE HW BDGP
Benchmarks: ALOI Caltech101-all NUS-WIDE HW BDGP
Metrics: ACC NMI ARI
Baselines: k-means MLAN MMGC MCGC CGL CVCL RCAGL SCM OMVCDR
Builds on: Shazeer et al. 2017 sparsely-gated mixture of experts
Closest papers by meaning
- Deep Variational Incomplete Multi-View Clustering: Exploring Shared Clustering Structures (AAAI 2024 · similarity 0.88)
- EASEMVC:Efficient Dual Selection Mechanism for Deep Multi-View Clustering (CVPR 2025 · similarity 0.88)
- Evaluate then Cooperate: Shapley-based View Cooperation Enhancement for Multi-view Clustering (NeurIPS 2024 · similarity 0.88)
- Investigating and Mitigating the Side Effects of Noisy Views for Self-Supervised Clustering Algorithms in Practical Multi-View Scenarios (CVPR 2024 · similarity 0.87)
- Differentiable Information Bottleneck for Deterministic Multi-view Clustering (CVPR 2024 · similarity 0.87)
- Deep Fair Multi-View Clustering with Attention KAN (CVPR 2025 · similarity 0.87)
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.