Research map / Label-efficient and robust learning / Recognition and long-tail
Deep Incomplete Multi-view Learning via Cyclic Permutation of VAEs
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 MVP, a multi-view VAE that models inter-view correspondences via permutations of latent variables and uses cyclic-permutation posteriors as informational priors to handle incomplete multi-view data.
Problem
Incomplete multi-view representation learning, where views are irregularly missing, yields representations lacking sufficiency and consistency, degrading clustering and generation as missing ratio rises.
Approach
Establishes inter-view latent correspondences by cross-view transformations, applies permutations and view-wise partitions to derive a valid ELBO, and uses cyclic permutations of posteriors as informational priors so the KL regularizer becomes a Permutation Divergence (a similarity measure across view distributions).
Key results
- On Handwritten at missing rate 0.1, MVP reaches 90.55% ACC, 87.08% NMI, 84.46% ARI, the best among compared methods.
- On CUB at missing rate 0.1, MVP achieves 78.67% ACC vs 72.93% for the second-best DSIMVC.
- On Reuters at missing rate 0.7, MVP leads the second-best method by 8.04% ACC; on PolyMNIST MVP achieves 83.56% semantic coherence accuracy when only view 2 is given.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MVP instead of CPSPAN | Incomplete multi-view clustering · Handwritten, missing rate 0.7 | ACC: 83.90 -> 86.74 | Helped |
| Use MVP instead of DSIMVC | Incomplete multi-view clustering · CUB, missing rate 0.7 | ACC: 67.33 -> 66.53 | Hurt |
| Use MVP instead of CPSPAN | Incomplete multi-view clustering · Scene15, missing rate 0.7 | ACC: 42.53 -> 43.14 | Helped |
| Use MVP instead of DSIMVC | Incomplete multi-view clustering · Reuters, missing rate 0.7 | ACC: 47.12 -> 55.16 | Helped |
| Use MVP instead of DSIMVC | Incomplete multi-view clustering · SensIT Vehicle, missing rate 0.7 | ACC: 66.54 -> 68.87 | Helped |
Surprising findings
- MMVAE and mmJSD slightly improve in semantic coherence and SSIM as more views are missing, attributed to MoE limitations in aggregating information.
Limitations
- On CUB with few views and small sample size, performance declines faster as missing ratio rises.
- Comparisons with MVAEs use the same network structure for all models rather than original architectures (e.g., MMVAE+ ResNet).
Experimental setup
Datasets: Handwritten CUB Scene15 Reuters SensIT Vehicle PolyMNIST MVShapeNet
Benchmarks: Handwritten CUB Scene15 Reuters SensIT Vehicle PolyMNIST MVShapeNet
Metrics: ACC NMI ARI SSIM coherence accuracy
Baselines: DCCA DCCAE DIMVC DSIMVC Completer CPSPAN ICMVC DVIMC mVAE mmVAE MoPoE mmJSD
Builds on: MVAE MMVAE MoPoE mmJSD MVTCAE MMVAE+
Closest papers by meaning
- Deep Variational Incomplete Multi-View Clustering: Exploring Shared Clustering Structures (AAAI 2024 · similarity 0.90)
- Rethinking Multi-view Representation Learning via Distilled Disentangling (CVPR 2024 · similarity 0.87)
- Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding (CVPR 2025 · similarity 0.87)
- Multi-View Causal Representation Learning with Partial Observability (ICLR 2024 · similarity 0.87)
- Incomplete Multi-view Clustering via Hierarchical Semantic Alignment and Cooperative Completion (NeurIPS 2025 · similarity 0.86)
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.86)
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.