Research map / Learning theory and algorithms / Clustering
Deep Variational Incomplete Multi-View Clustering: Exploring Shared Clustering Structures
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 DVIMC, an imputation-free deep variational method for incomplete multi-view clustering that fuses view-specific VAE encoders via Product-of-Experts with a coherence KL loss and a VaDE-style Mixture-of-Gaussians prior.
Problem
Incomplete multi-view clustering, where samples miss some views; imputation-based methods risk inaccurate imputations and imputation-free methods suffer from information imbalance across views and under-exploit shared information.
Approach
Aggregates available view-specific VAE posteriors with Product-of-Experts into a common latent, adds a coherence loss (average KL between the aggregated and each single-view posterior) to counter information imbalance, and imposes a Mixture-of-Gaussians prior with the VaDE trick so the shared representation is clustering-friendly.
Key results
- On Caltech7-5V at missing rate 0.5, DVIMC's ACC and NMI are about 10% and 15% higher than the second-best method DIMVC.
- On NoisyMNIST at missing rate 0.7, DVIMC reaches ACC 0.9118 versus 0.5443 for MVAE, roughly a 37% improvement.
- DVIMC outperforms the other VAE-based methods (MVAE, DMVCVAE, Multi-VAE) on all four datasets across all metrics in the reported tables.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the coherence objective loss to DVIMC (vanilla) (ablation) | DVIMC · Four datasets at different missing view rates | clustering performance | Helped |
| Use Product-of-Experts aggregation instead of weight summarization (ablation) | DVIMC (vanilla) · Four datasets at different missing view rates | clustering performance | Helped |
| Use a clustering-structured latent prior and VaDE trick instead of MVAE's latent prior (ablation) | DVIMC (vanilla) · Four datasets at different missing view rates | clustering performance | Helped |
| Use DVIMC instead of the strongest baseline | Caltech7-5V · Caltech7-5V, missing rates 0.1, 0.3, 0.5, and 0.7 | ACC: 0.895 vs 0.834; 0.886 vs 0.798; 0.868 vs 0.760; 0.844 vs 0.715 | Helped |
| Use DVIMC instead of the strongest baseline | Scene-15 · Scene-15, missing rates 0.1, 0.3, 0.5, and 0.7 | ACC: 0.479 vs 0.456; 0.454 vs 0.424; 0.442 vs 0.412; 0.411 vs 0.386 | Helped |
| Use DVIMC instead of the strongest baseline | Multi-Fashion · Multi-Fashion, missing rates 0.1, 0.3, 0.5, and 0.7 | ACC: 0.883 vs 0.880; 0.879 vs 0.873; 0.824 vs 0.835; 0.815 vs 0.796 | Mixed |
| Use DVIMC instead of the strongest baseline | NoisyMNIST · NoisyMNIST, missing rates 0.1, 0.3, 0.5, and 0.7 | ACC: 0.965 vs 0.891; 0.943 vs 0.912; 0.923 vs 0.894; 0.911 vs 0.875 | Helped |
| Increase the missing view rate | DVIMC · Four datasets | ACC: 0.895 to 0.844; 0.479 to 0.411; 0.883 to 0.815; 0.965 to 0.911 | Hurt |
Surprising findings
- The paper reports that the VAE-based baselines degrade sharply as the missing rate grows, while the imputation-based DCP is strong on NoisyMNIST at some rates.
Limitations
- Regularization parameter alpha must be tuned per dataset (5, 10, or 20 used), and too large or small values hurt clustering.
- Imputation-free design relies on the PoE aggregation, which may still be dominated by low-variance views in some regimes.
- Evaluation is limited to four datasets with synthetic random missing-view construction.
Experimental setup
Datasets: Caltech7-5V Scene-15 Multi-Fashion NoisyMNIST
Benchmarks: Caltech7-5V Scene-15 Multi-Fashion NoisyMNIST
Metrics: ACC NMI ARI PUR
Baselines: MVAE DMVCVAE Multi-VAE DIMVC DCP DSIMVC APADC
Builds on: VaDE Product-of-Experts MVAE DMVCVAE
Closest papers by meaning
- Deep Incomplete Multi-View Clustering via Hierarchical Imputation and Alignment (AAAI 2026 · similarity 0.95)
- Geometry-Aware Variational Information Maximization for Deep Incomplete Multi-view Clustering (AAAI 2026 · similarity 0.94)
- Adaptive Feature Imputation with Latent Graph for Deep Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.94)
- Towards Resource-friendly, Extensible and Stable Incomplete Multi-view Clustering (ICML 2024 · similarity 0.92)
- Incomplete Multi-view Clustering via Diffusion Contrastive Generation (AAAI 2025 · similarity 0.92)
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.92)
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.