Research map / Learning theory and algorithms / Clustering
Incomplete Multi-view Clustering via Diffusion Contrastive Generation
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 Diffusion Contrastive Generation (DCG), an end-to-end incomplete multi-view clustering method that uses intra-view diffusion plus contrastive alignment on few paired samples to recover missing views, with instance- and category-level interactive learning.
Problem
Incomplete multi-view clustering suffers when recovering missing views relies heavily on paired data (impractical at high missing rates) and generated views lack diversity and discriminability.
Approach
Applies forward diffusion and reverse denoising to intra-view latent representations, then uses contrastive learning on a small set of paired samples to align generated views with real views; exploits the observation that denoising pulls samples toward cluster centers, so extrapolating the diffusion step at inference improves clustering.
Key results
- On CUB at missing rate 0.1, DCG reaches 82.23 ACC, 77.70 NMI, 69.21 ARI, beating the second-best method by 5.56%, 6.32%, and 10.56% respectively.
- At missing rate 0.7 on CUB, DCG outperforms state-of-the-art by 6.67% per the paper's introduction.
- On HandWritten, DCG's ACC drops only 3% from missing rate 0.1 to 0.7, versus 8.88% for ICMVC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DCG instead of the strongest listed baseline | Clustering methods · Synthetic3d, missing rate 0.3 | ACC: 87.50 -> 88.00 | Helped |
| use DCG instead of the strongest listed baseline | Clustering methods · CUB, missing rate 0.3 | ARI: 56.84 -> 59.85 | Helped |
| use DCG instead of the strongest listed baseline | Clustering methods · HandWritten, missing rate 0.3 | ACC: 82.27 -> 82.70 | Helped |
| use DCG instead of the strongest listed baseline | Clustering methods · LandUse-21, missing rate 0.3 | NMI: 30.18 -> 32.09 | Helped |
| use DCG instead of the strongest listed baseline | Clustering methods · Fashion, missing rate 0.3 | ACC: 74.11 -> 93.13 | Helped |
| add diffusion loss (ablation) | DCG · CUB, missing rate 0.3 | ACC: 70.96 -> 77.17 | Helped |
| add generation contrastive loss (ablation) | DCG · CUB, missing rate 0.3 | ACC: 71.46 -> 77.17 | Helped |
| add diffusion and generation contrastive losses (ablation) | DCG · CUB, missing rate 0.3 | ACC: 60.73 -> 77.17 | Helped |
| add maximum mutual information loss (ablation) | DCG · CUB, missing rate 0.3 | ACC: 72.83 -> 77.17 | Helped |
| add category-level contrastive loss (ablation) | DCG · CUB, missing rate 0.3 | ACC: 71.23 -> 77.17 | Helped |
| add KL divergence loss (ablation) | DCG · CUB, missing rate 0.3 | ACC: 73.63 -> 77.17 | Helped |
| add category-level contrastive and KL divergence losses (ablation) | DCG · CUB, missing rate 0.3 | ACC: 68.89 -> 77.17 | Helped |
Surprising findings
- Extrapolating the diffusion step at inference (from 50 to 100) increases clustering accuracy, linking diffusion denoising to cluster compactness.
Limitations
- Relies on a limited set of paired multi-view samples for cross-view alignment.
- Hyperparameter sensitivity analysis is reported only at missing rate 0.3 on selected datasets.
Experimental setup
Datasets: Synthetic3D CUB HandWritten LandUse-21 Fashion
Benchmarks: Synthetic3D CUB HandWritten LandUse-21 Fashion
Metrics: ACC NMI ARI
Baselines: DCP DSIMVC GCFAgg CPSPAN APADC ProImp DVIMVC ICMVC
Builds on: denoising diffusion probabilistic models contrastive learning maximum mutual information multi-view clustering category-level contrastive clustering
Closest papers by meaning
- Diffusion-based Missing-view Generation With the Application on Incomplete Multi-view Clustering (ICML 2024 · similarity 0.94)
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.93)
- Adaptive Feature Imputation with Latent Graph for Deep Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.92)
- Deep Variational Incomplete Multi-View Clustering: Exploring Shared Clustering Structures (AAAI 2024 · similarity 0.92)
- Towards Resource-friendly, Extensible and Stable Incomplete Multi-view Clustering (ICML 2024 · similarity 0.91)
- Medusa: A Multi-Scale High-order Contrastive Dual-Diffusion Approach for Multi-View Clustering (CVPR 2025 · similarity 0.90)
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.