Research map / Federated learning and privacy / Personalized and heterogeneous FL
Bridging Gaps: Federated Multi-View Clustering in Heterogeneous Hybrid Views
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 FMCSC, a federated multi-view clustering method for heterogeneous hybrid views (mixed single-view and multi-view clients) using cross-client consensus pre-training, local-synergistic contrastive learning, and global-specific weighting aggregation, evaluated on four multi-view datasets against centralized incomplete MVC and federated baselines.
Problem
Federated multi-view clustering when clients are heterogeneous, with a mixture of single-view and multi-view clients (client gap and view gap), where prior FedMVC methods assume isomorphic clients.
Approach
Uses contrastive learning at feature level for multi-view clients and model level (against global model) for single-view clients, with server aggregation weights derived from a mutual-information bound, plus a consensus pre-training step to avoid model misalignment.
Key results
- On MNIST-USPS with M/S=1:1, FMCSC reaches 92.9 ACC / 84.2 NMI / 85.0 ARI, versus 87.3 ACC for the next best (ProImp).
- On Multi-Fashion with M/S=2:1, FMCSC reaches 92.4 ACC vs 82.7 for DSIMVC (second best).
- On NUSWIDE with M/S=1:1, FMCSC reaches 52.7 ACC vs 53.0 for JPLTD at M/S=2:1 region; at M/S=1:1 FMCSC 52.7 ACC is best vs 51.5 for JPLTD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FMCSC instead of the strongest baseline | Clustering on MNIST-USPS, BDGP, Multi-Fashion, and NUSWIDE at M/S = 1:1 · MNIST-USPS; BDGP; Multi-Fashion; NUSWIDE | ACC: MNIST-USPS: 87.3±0.6 -> 92.9±1.2; BDGP: 91.5±0.5 -> 91.9±1.2; Multi-Fashion: 77.7±1.4 -> 90.4±0.6; NUSWIDE: 51.5±0.5 -> | Helped |
| Add consensus pre-training (ablation) | FMCSC clustering · MNIST-USPS, BDGP, Multi-Fashion, and NUSWIDE, M/S = 1:1 | ACC: MNIST-USPS: 91.12 -> 93.38; BDGP: 69.84 -> 91.92; Multi-Fashion: 87.80 -> 90.36; NUSWIDE: 42.16 -> 54.04 | Helped |
| Add multi-view client global-model alignment (ablation) | FMCSC clustering · MNIST-USPS, BDGP, Multi-Fashion, and NUSWIDE, M/S = 1:1 | ACC: MNIST-USPS: 60.82 -> 93.38; BDGP: 61.04 -> 91.92; Multi-Fashion: 56.04 -> 90.36; NUSWIDE: 42.08 -> 54.04 | Helped |
| Add single-view client model comparison (ablation) | FMCSC clustering · MNIST-USPS, BDGP, Multi-Fashion, and NUSWIDE, M/S = 1:1 | ACC: MNIST-USPS: 88.66 -> 93.38; BDGP: 87.40 -> 91.92; Multi-Fashion: 82.58 -> 90.36; NUSWIDE: 45.86 -> 54.04 | Helped |
| Add weighted aggregation (ablation) | FMCSC clustering · MNIST-USPS, BDGP, Multi-Fashion, and NUSWIDE, M/S = 1:1 | ACC: MNIST-USPS: 89.52 -> 93.38; BDGP: 85.44 -> 91.92; Multi-Fashion: 88.14 -> 90.36; NUSWIDE: 47.54 -> 54.04 | Helped |
| Set contrastive temperatures to 0.1 rather than the tested higher values (ablation) | FMCSC on MNIST-USPS · MNIST-USPS | clustering performance | Hurt |
| Increase the proportion of single-view clients (ablation) | FMCSC and comparison methods · Heterogeneous hybrid-view scenarios | clustering performance | Hurt |
| Add differential-privacy noise at ε = 10 rather than ε = 50 (ablation) | FMCSC · Privacy evaluation | clustering accuracy | Hurt |
Surprising findings
- Removing consensus pre-training (Item-1) on MNIST-USPS and Multi-Fashion still reached reasonable results only after running four times more communication rounds, showing initial misalignment matters but can be partly mitigated by more rounds.
Limitations
- Performance declines as the number of clients grows to 100 due to fewer samples per client.
- Generalization bound depends on view-distribution divergence, which remains high for dissimilar views.
- Evaluated only on four multi-view datasets with simulated client partitions.
Experimental setup
Datasets: MNIST-USPS BDGP Multi-Fashion NUSWIDE
Benchmarks: MNIST-USPS BDGP Multi-Fashion NUSWIDE
Metrics: ACC NMI ARI
Baselines: HCP-IMSC IMVC-CBG DSIMVC LSIMVC ProImp JPLTD CPSPAN FedDMVC FCUIF
Builds on: federated learning (FedAvg-style aggregation) contrastive learning (feature contrast) MoCo-style model contrast autoencoder pre-training K-means
Closest papers by meaning
- Graph Consistency and Diversity Measurement for Federated Multi-View Clustering (AAAI 2025 · similarity 0.91)
- Scalable Federated One-Step Multi-View Clustering with Tensorized Regularization (AAAI 2025 · similarity 0.91)
- Federated Incomplete Multi-view Clustering with Globally Fused Graph Guidance (ICML 2025 · similarity 0.89)
- Efficient Federated Incomplete Multi-View Clustering (ICML 2025 · similarity 0.89)
- EASEMVC:Efficient Dual Selection Mechanism for Deep Multi-View Clustering (CVPR 2025 · similarity 0.88)
- An Effective and Secure Federated Multi-View Clustering Method with Information-Theoretic Perspective (ICML 2025 · similarity 0.87)
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.