Research map / Continual learning / Domain and prompt-based
AdaptCMVC: Robust Adaption to Incremental Views in Continual 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.
Reframes continual multi-view clustering as domain adaptation: a VAE base model is adapted to each incoming view via a weight-averaged teacher self-training loop with noise-weighted consistency and a structure alignment loss.
Problem
Continual multi-view clustering, where views arrive sequentially, past-view data is inaccessible, and models must handle view discrepancy, view-specific noise, and catastrophic forgetting.
Approach
Instead of learning a separate model per view and fusing partitions by moving average (late fusion), adapt a single base VAE to each new view with an EMA-teacher self-training loop whose consistency loss is down-weighted for samples far from last-session K-means prototypes, plus a similarity-matrix structure alignment term that stores cross-view group structure without accessing old data.
Key results
- On E-MNIST, AdaptCMVC reaches 67.10 ACC / 54.14 NMI versus 57.30 ACC for the best CMVC baseline (CMVC).
- On PatchedMNIST, AdaptCMVC achieves 84.33 ACC and 55.87 NMI, exceeding CMVC/CAC by about 13 ACC points.
- Averaged over all datasets, AdaptCMVC attains 58.30 ACC and 56.23 NMI with the best average rank (2.3 ACC, 3.5 NMI).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AdaptCMVC instead of the strongest reported baseline | continual multi-view clustering · E-MNIST | ACC: 67.10 -> 67.10 | No clear effect |
| Use AdaptCMVC instead of the strongest reported baseline | continual multi-view clustering · E-FMNIST | ACC: 53.16 -> 54.13 | Helped |
| Use AdaptCMVC instead of the strongest reported baseline | continual multi-view clustering · Office-31 | ACC: 33.16 -> 22.81 | Hurt |
| Use AdaptCMVC instead of the strongest reported baseline | continual multi-view clustering · COIL-100 | ACC: 60.31 -> 57.12 | Hurt |
| Use AdaptCMVC instead of the strongest reported baseline | continual multi-view clustering · COIL-20 | ACC: 68.50 -> 64.32 | Hurt |
| Use AdaptCMVC instead of the strongest reported baseline | continual multi-view clustering · PatchedMNIST | ACC: 83.20 -> 84.33 | Helped |
| add the noise-robust consistency loss (ablation) | AdaptCMVC · E-MNIST, Office-31, COIL-100, and COIL-20 | clustering performance | Helped |
| add the structure alignment mechanism (ablation) | AdaptCMVC · E-MNIST, Office-31, COIL-100, and COIL-20 | clustering performance | Helped |
Surprising findings
- On COIL-20 AdaptCMVC underperforms CMVC and CAC in ACC (64.32 vs 68.15/68.50), and on the Office-31 and COIL-100 datasets the advantage is inconsistent across metrics.
Limitations
- Performance on COIL-20 is lower than the CMVC and CAC baselines in ACC.
- Requires a base model trained on the first view and per-session K-means prototypes from the previous session, adding dependence on the earlier clustering.
- Evaluated only on image multi-view datasets with identical feature dimensions across views.
Experimental setup
Datasets: E-MNIST E-FMNIST Office-31 COIL-100 COIL-20 PatchedMNIST
Benchmarks: E-MNIST E-FMNIST Office-31 COIL-100 COIL-20 PatchedMNIST
Metrics: ACC NMI
Baselines: Joint-VAE β-VAE MFLVC CONAN EAMC GCFAgg Multi-VAE CMVC CAC
Builds on: VAE mean teacher continual domain adaptation self-training CMVC CAC
Closest papers by meaning
- Robust Multi-view Clustering against Imperfect Information (NeurIPS 2026 · 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 0.87)
- Bridging Gaps: Federated Multi-View Clustering in Heterogeneous Hybrid Views (NeurIPS 2024 · similarity 0.87)
- Live and Learn: Continual Action Clustering with Incremental Views (AAAI 2024 · similarity 0.86)
- Investigating and Mitigating the Side Effects of Noisy Views for Self-Supervised Clustering Algorithms in Practical Multi-View Scenarios (CVPR 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.