Research map / Learning theory and algorithms / Clustering
Enhanced then Progressive Fusion with View Graph 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 EPFMVC, a deep multi-view clustering framework combining SE-based channel-attention feature enhancement with OT-distance view-graph progressive fusion and contrastive intra-view and cross-view losses.
Problem
Deep multi-view clustering suffers from feature conflicts when directly fusing views with divergent distributions and from insufficient enhancement of per-view feature channels.
Approach
Enhance each view's feature channels with an SE-augmented multi-head self-attention encoder, then build a view graph from pairwise optimal-transport distances and progressively fuse each view with its nearest view via cross-view attention, aligned to the shared representation by contrastive losses.
Key results
- On BBC, UCI-digit and CIFAR10-All, EPFMVC improves over the second-best method by 11.97%, 3.25% and 4.5% respectively (as stated by the authors).
- On NGs, EPFMVC reaches 95.00 ACC and 84.97 NMI versus 94.00 ACC for the best baseline SEM.
- Progressive fusion beats direct concatenation fusion by 19.86%, 9.35% and 3.04% ACC on BBC, HW and CIFAR10-All.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EPFMVC instead of BMVC | Multi-view clustering · BBC | ACC (%): 11.97% | Helped |
| Use EPFMVC instead of MFLVC | Multi-view clustering · UCI-digit | ACC (%): 3.25% | Helped |
| Use EPFMVC instead of MVCAN | Multi-view clustering · CIFAR10-All | ACC (%): 4.5% | Helped |
| Replace the autoencoder with FCAencoder (ablation) | EPFMVC · NGs | ACC (%): 47.20 -> 95.00 | Helped |
| Replace the autoencoder with FCAencoder (ablation) | EPFMVC · BBC | ACC (%): 80.58 -> 87.01 | Helped |
| Replace the autoencoder with FCAencoder (ablation) | EPFMVC · UCI-digit | ACC (%): 93.55 -> 95.25 | Helped |
| Replace the autoencoder with FCAencoder (ablation) | EPFMVC · HW | ACC (%): 81.65 -> 96.90 | Helped |
| Replace the autoencoder with FCAencoder (ablation) | EPFMVC · Hdigit | ACC (%): 99.46 -> 99.88 | Helped |
| Replace the autoencoder with FCAencoder (ablation) | EPFMVC · Wiki | ACC (%): 59.39 -> 63.19 | Helped |
| Replace the autoencoder with FCAencoder (ablation) | EPFMVC · CIFAR10-All | ACC (%): 30.65 -> 36.94 | Helped |
| Replace direct concatenation fusion with progressive view-graph fusion (ablation) | EPFMVC · BBC | ACC: 67.15 -> 87.01 | Helped |
Surprising findings
- Replacing the standard autoencoder with the FCAencoder gives a 47.8% ACC gain on NGs, far larger than other reported gains.
Limitations
- Fusion only pairs each view with its single nearest view, a heuristic that may not capture global structure.
- Evaluation limited to seven benchmark datasets and nine baselines, with no comparison to newer methods beyond 2024.
Experimental setup
Datasets: NGs BBC HandWritten UCI-digit Hdigit Wiki CIFAR10-All
Benchmarks: NGs BBC HandWritten UCI-digit Hdigit Wiki CIFAR10-All
Metrics: ACC NMI Purity F-score
Baselines: BMVC LMVSC AE2-Nets AdaGAE SDMVC MFLVC DSMVC SEM MVCAN
Builds on: Squeeze-and-Excitation multi-head self-attention optimal transport contrastive learning
Closest papers by meaning
- Incomplete and Unpaired Multi-View Graph Clustering with Cross-View Feature Fusion (AAAI 2025 · similarity 0.89)
- Sharper Error Bounds in Late Fusion Multi-view Clustering with Eigenvalue Proportion Optimization (AAAI 2025 · similarity 0.89)
- EASEMVC:Efficient Dual Selection Mechanism for Deep Multi-View Clustering (CVPR 2025 · similarity 0.88)
- Medusa: A Multi-Scale High-order Contrastive Dual-Diffusion Approach for Multi-View Clustering (CVPR 2025 · similarity 0.88)
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.88)
- Deep Fair Multi-View Clustering with Attention KAN (CVPR 2025 · 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.