Research map / Graph learning / Graph self-supervised learning
EASEMVC:Efficient Dual Selection Mechanism 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 EASEMVC, a deep multi-view clustering framework that selects view pairs via an OT-based view graph and reweights samples via a cross-view sample graph before contrastive learning and DDC clustering loss.
Problem
View-pair contrastive learning in multi-view clustering degrades when views are low-quality or non-redundant, and existing methods pair views arbitrarily or include all views without prior knowledge of which pairs are suitable.
Approach
Builds a view graph from optimal-transport distances between anchor-based bipartite graphs to pick the nearest view for each view, and a sample graph from OT distances to weight positive sample pairs in a weighted contrastive loss, jointly optimized with reconstruction and deep divergence clustering losses.
Key results
- On ALOI-100, EASEMVC reaches 83.81% ACC, about 20 points above the second-best method, per the paper.
- On HandWritten, EASEMVC reports 96.05% ACC and 92.67% NMI, the best among compared methods in Table 1.
- Ablation on ALOI-100 shows the full model (83.81 ACC) improves over reconstruction-only (70.59 ACC).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add contrastive learning loss to reconstruction-only training (ablation) | EASEMVC · ALOI-100 | ACC: 70.59 -> 75.79 | Helped |
| Add DDC clustering loss to reconstruction plus contrastive learning (ablation) | EASEMVC · ALOI-100 | ACC: 75.79 -> 83.81 | Helped |
| Add DDC clustering loss to reconstruction-only training (ablation) | EASEMVC · ALOI-100 | ACC: 70.59 -> 78.98 | Helped |
| Add contrastive learning loss to reconstruction-only training (ablation) | EASEMVC · BBC | ACC: 54.45 -> 78.25 | Helped |
| Add DDC clustering loss to reconstruction plus contrastive learning (ablation) | EASEMVC · BBC | ACC: 78.25 -> 83.94 | Helped |
| Add DDC clustering loss to reconstruction-only training (ablation) | EASEMVC · BBC | ACC: 54.45 -> 57.96 | Helped |
| Use adaptive view selection instead of contrastive learning uniformly across views (ablation) | EASEMVC · ALOI-100 and BBC | clustering performance | Helped |
| Use selective cross-view consistency with sample weights instead of removing positive-pair weights (ablation) | EASEMVC · ALOI-100 and BBC | clustering performance | Helped |
| Use adaptive view selection on the view topology graph (ablation) | EASEMVC · ALOI-100 and BBC | clustering performance | Helped |
Surprising findings
- Forcing contrastive alignment across all views degraded the quality of high-quality views; e.g., Caltech-5V fused accuracy dropped from 84.57% to 73.79% after multi-view contrastive learning.
Limitations
- Theoretical justification of the non-transitive redundancy assumption is deferred to the supplementary material.
- Text datasets (BBC, FashionMNIST) show second-best rather than best results on several metrics.
Experimental setup
Datasets: Caltech101-20 HandWritten UCI ALOI-100 BBC FashionMNIST
Benchmarks: Caltech101-20 HandWritten UCI ALOI-100 BBC FashionMNIST
Metrics: ACC NMI Purity F-score
Baselines: LMVSC BMVC AE2-Nets SDMVC MFLVC DSMVC DEMVC SDSNE SEM SCMVC
Builds on: DDC (deep divergence clustering) CoMVC DCP InfoNCE
Closest papers by meaning
- Medusa: A Multi-Scale High-order Contrastive Dual-Diffusion Approach for Multi-View Clustering (CVPR 2025 · similarity 0.92)
- Dual-stage Contrastive Learning-enhanced Multi-view Variational Clustering (ICML 2026 · similarity 0.91)
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.90)
- SURER: Structure-Adaptive Unified Graph Neural Network for Multi-View Clustering (AAAI 2024 · similarity 0.89)
- Deep Fair Multi-View Clustering with Attention KAN (CVPR 2025 · similarity 0.89)
- Multi-View Graph Clustering via Node-Guided Contrastive Encoding (ICML 2025 · 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.