Research map / Federated learning and privacy / Personalized and heterogeneous FL
FedMSGL: A Self-Expressive Hypergraph Based Federated Multi-View Learning
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 FedMSGL, a vertical federated multi-view clustering method that uses self-expressive subspace learning to get uniform-dimension subspaces locally and builds a global hypergraph from consistent and view-specific subspaces at the server.
Problem
Vertical federated learning overlooks differences in feature dimensions across participants (biasing global models toward high-dimensional nodes) and does not handle multi-view data well, so federated multi-view clustering remains underexplored.
Approach
Local nodes learn self-expressive subspace embeddings of equal dimension (n x n), split into consistent and view-specific parts with a manifold coefficient matrix; the server fuses consistent parts with adaptive inverse-distance weights and builds a global k-NN hypergraph from the global and view-specific affinities to get the clustering indicator via spectral decomposition.
Key results
- On BBC Sport, FedMSGL reaches ACC 0.9022, Purity 0.9061 and NMI 0.7543, outperforming the best centralized baseline FSMSC (ACC 0.8272).
- On ORL, FedMSGL achieves ACC 0.8283 versus 0.8025 for FSMSC; on Handwritten ACC 0.8898 versus FSMSC 0.8845.
- On Caltech101-7, FedMSGL (ACC 0.7256) trails CDMGC (ACC 0.7631) and on Sonar NMI and ACC FedMVL is higher in some cases, so gains are not uniform across datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedMSGL instead of FedMVL | federated multi-view clustering · Sonar | ACC: 0.6250±0.00 -> 0.6466±0.00 | Helped |
| Use FedMSGL instead of FSMSC | multi-view clustering · BBC Sport | ACC: 0.8272±0.00 -> 0.9022±0.04 | Helped |
| Use FedMSGL instead of FSMSC | multi-view clustering · ORL | ACC: 0.8025±0.00 -> 0.8283±0.02 | Helped |
| Use FedMSGL instead of FSMSC | multi-view clustering · Handwritten | ACC: 0.8845±0.00 -> 0.8898±0.08 | Helped |
| Use FedMSGL instead of CDMGC | multi-view clustering · Caltech101-7 | ACC: 0.7631±0.05 -> 0.7256±0.07 | Hurt |
| Use FedMSGL instead of SFMC | multi-view clustering · Sonar | Purity: 0.6826±0.00 -> 0.6466±0.00 | Hurt |
| Use FedMSGL instead of FSMSC | multi-view clustering · BBC Sport | Purity: 0.8327±0.00 -> 0.9061±0.02 | Helped |
| Use FedMSGL instead of CDMGC | multi-view clustering · ORL | Purity: 0.8600±0.01 -> 0.8520±0.01 | Hurt |
| Use FedMSGL instead of FSMSC | multi-view clustering · Handwritten | Purity: 0.8845±0.00 -> 0.8964±0.08 | Helped |
| Use FedMSGL instead of CDMGC | multi-view clustering · Caltech101-7 | Purity: 0.8937±0.09 -> 0.8453±0.01 | Hurt |
| Use FedMSGL instead of FedMVL | federated multi-view clustering · Sonar | NMI: 0.1398±0.00 -> 0.1204±0.00 | Hurt |
| Use FedMSGL instead of FSMSC | multi-view clustering · BBC Sport | NMI: 0.7217±0.00 -> 0.7543±0.03 | Helped |
Surprising findings
- On Sonar and Caltech101-7 FedMSGL does not lead on all metrics; e.g., FedMVL has higher ACC on Sonar (0.6250 vs 0.6466 is FedMSGL higher, but NMI on Sonar is higher for FedMVL, 0.1398 vs 0.1204).
Limitations
- Performance is sensitive to hyperparameters lambda_1 and lambda_3 on BBC Sport.
- Gains are not uniform across all datasets and metrics (e.g., Caltech101-7 and Sonar NMI).
- Experiments are clustering-only on five small multi-view datasets; the appendix containing details was not provided.
Experimental setup
Datasets: Sonar BBC Sport ORL Handwritten Caltech101-7
Benchmarks: Sonar BBC Sport ORL Handwritten Caltech101-7
Metrics: ACC Purity NMI
Baselines: SC SFMC CDMGC FSMSC RCAGL FedMVL FedMVFCM FedMVFPC
Builds on: self-expressive subspace learning hypergraph Laplacian FedBCD FedV
Closest papers by meaning
- FedNE: Surrogate-Assisted Federated Neighbor Embedding for Dimensionality Reduction (NeurIPS 2024 · similarity 0.87)
- FedWon: Triumphing Multi-domain Federated Learning Without Normalization (ICLR 2024 · similarity 0.87)
- Federated Incomplete Multi-view Clustering with Globally Fused Graph Guidance (ICML 2025 · similarity 0.86)
- Graph Consistency and Diversity Measurement for Federated Multi-View Clustering (AAAI 2025 · similarity 0.86)
- FedLMT: Tackling System Heterogeneity of Federated Learning via Low-Rank Model Training with Theoretical Guarantees (ICML 2024 · similarity 0.86)
- An Effective and Secure Federated Multi-View Clustering Method with Information-Theoretic Perspective (ICML 2025 · 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.