Research map / Learning theory and algorithms / Clustering
Deep Streaming 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 Deep Streaming View Clustering (DSVC), an autoencoder-based method with a knowledge base that aligns prototype knowledge across arriving views to mitigate concept drift in streaming multi-view clustering.
Problem
Multi-view clustering assumes all views are available at once, but in practice views arrive as a stream with distribution shifts (concept drift) across views, which degrades existing lifelong/static methods.
Approach
A knowledge base of prototype knowledge is carried across views; each new view's prototypes are aligned to historical prototypes via a symmetric KL distribution-consistency loss, and used to guide feature clustering structure, with only the current view processed at each step.
Key results
- On ALOI-10, DSVC reaches 86.69% ACC versus 82.33% for CVCL and 85.64% for GDMVC (suboptimal), a 1.14% gain on ACC as reported in the text.
- On Stl10-fea, DSVC achieves 63.41% ACC vs. 56.37% for MAGA, while many static DMVC methods run out of memory on ALOI-1000 and YoutubeFace.
- On HandWritten, DSVC reaches 95.44% ACC, above the best baseline MVCAN at 94.12%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DSVC instead of DMCE | streaming view clustering · HandWritten | ACC: 95.44±0.32 vs 94.30±3.97 | Helped |
| use DSVC instead of GDMVC | streaming view clustering · ALOI-10 | ACC: 86.69±3.85 vs 85.64±4.08 | Helped |
| use DSVC instead of DCP | streaming view clustering · LandUse-21 | ACC: 28.52±1.16 vs 28.29±1.36 | Helped |
| use DSVC instead of DSRL | streaming view clustering · Scene-15 | ACC: 45.33±0.88 vs 45.53±0.86 | Hurt |
| use DSVC instead of MAGA | streaming view clustering · Stl10-fea | ACC: 63.41±9.96 vs 56.37±16.12 | Helped |
| use DSVC instead of DMCE | streaming view clustering · ALOI-100 | ACC: 80.53±0.59 vs 74.83±0.52 | Helped |
| use DSVC instead of MAGA | streaming view clustering · YoutubeFace | ACC: 25.09±1.09 vs 22.95±1.36 | Helped |
| use DSVC instead of DCP | streaming view clustering · ALOI-1000 | ACC: 58.32±0.56 vs 48.12±0.86 | Helped |
| add knowledge guidance learning (ablation) | DSVC · ALOI-10 | ACC: 74.42 -> 86.69 | Helped |
| add knowledge guidance learning (ablation) | DSVC · HandWritten | ACC: 86.45 -> 95.44 | Helped |
| add knowledge guidance learning (ablation) | DSVC · LandUse-21 | ACC: 24.86 -> 28.52 | Helped |
| add knowledge guidance learning (ablation) | DSVC · Scene-15 | ACC: 39.60 -> 45.33 | Helped |
Surprising findings
- Several static DMVC methods (e.g., DSRL, COMIC on large datasets) run out of memory on the larger datasets, while the streaming method scales to ALOI-1000.
Limitations
- Concept drift is handled via a single prototype-alignment loss; robustness to very different view orders is only partially addressed (performance gap across sequences shrinks but remains early on).
- Several hyperparameters (alpha, beta, M) are tuned per dataset, with different settings for some datasets.
Experimental setup
Datasets: HandWritten ALOI-10 LandUse-21 Scene-15 ALOI-100 Stl10-fea YoutubeFace ALOI-1000
Benchmarks: HandWritten ALOI-10 LandUse-21 Scene-15 ALOI-100 Stl10-fea YoutubeFace ALOI-1000
Metrics: ACC NMI ARI
Baselines: COMIC DSRL MFLVC DCP CVCL DealMVC CPSPAN DMCE GDMVC MAGA FMCSC MVCAN
Builds on: LSVC lifelong learning
Closest papers by meaning
- Mixture of Experts as Representation Learner for Deep Multi-View Clustering (AAAI 2025 · similarity 0.86)
- Deep Fair Multi-View Clustering with Attention KAN (CVPR 2025 · similarity 0.86)
- Towards Resource-friendly, Extensible and Stable Incomplete Multi-view Clustering (ICML 2024 · similarity 0.85)
- AdaptCMVC: Robust Adaption to Incremental Views in Continual Multi-view Clustering (CVPR 2025 · similarity 0.85)
- Enhanced then Progressive Fusion with View Graph for Multi-View Clustering (CVPR 2025 · similarity 0.85)
- Investigating and Mitigating the Side Effects of Noisy Views for Self-Supervised Clustering Algorithms in Practical Multi-View Scenarios (CVPR 2024 · similarity 0.85)
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.