Research map / Federated learning and privacy / Personalized and heterogeneous FL
Graph Consistency and Diversity Measurement for Federated 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.
Proposes MGCD, a graph-based federated multi-view clustering framework where clients upload view-specific diversity graphs (not features) and the server iteratively fuses them into a global consistency graph via contrastive fusion and diversity penalties.
Problem
Federated multi-view clustering must fuse heterogeneous views held on separate devices while preserving privacy; prior methods upload latent features vulnerable to model inversion and ignore view-specific diversity.
Approach
Clients upload only sample-similarity (diversity) graphs produced by a dual autoencoder plus self-representation, and the server distills residual consistency from these graphs and feeds the previous-round global consistency graph back in, iterating so consistency accumulates; a diversity penalty makes the consistency graph more compact.
Key results
- On Animal, MGCD reaches 0.5201 ACC and 0.6491 NMI versus the best baseline's roughly 0.19 ACC and 0.17 NMI, i.e. ~32.91% and ~48.31% gains over the sub-optimal method as reported.
- On NoisyMNIST, MGCD achieves 0.7791 ACC and 0.7261 NMI, outperforming the strongest baseline CoMVC (0.6465 ACC, 0.6469 NMI).
- Ablation on Aloi shows the full model reaches 0.916 ACC and 0.899 ARI, above removing graph consistency or diversity measurement terms.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MGCD graph consistency and diversity measurements | federated multi-view clustering · Mfeat | ACC: 0.9365 -> 0.9390 | Helped |
| use MGCD graph consistency and diversity measurements | federated multi-view clustering · Scene | ACC: 0.4383 -> 0.5285 | Helped |
| use MGCD graph consistency and diversity measurements | federated multi-view clustering · Aloi | ACC: 0.8566 -> 0.9160 | Helped |
| use MGCD graph consistency and diversity measurements | federated multi-view clustering · Animal | ACC: 0.1910 -> 0.5201 | Helped |
| use MGCD graph consistency and diversity measurements | federated multi-view clustering · Cifar10 | ACC: 0.9925 -> 0.9948 | Helped |
| use MGCD graph consistency and diversity measurements | federated multi-view clustering · NoisyMNIST | ACC: 0.6465 -> 0.7791 | Helped |
| add graph diversity measurement Ldd (ablation) | MGCD · Aloi | ACC: 0.853 -> 0.916 | Helped |
| add graph consistency measurement Lcl (ablation) | MGCD · Aloi | ACC: 0.903 -> 0.916 | Helped |
| add sample representation loss Lr (ablation) | MGCD · Aloi | ACC: 0.807 -> 0.916 | Helped |
| add graph consistency and diversity measurements Lcl and Ldd (ablation) | MGCD · Aloi | ACC: 0.795 -> 0.916 | Helped |
| use dual autoencoder instead of traditional autoencoder (ablation) | MGCD client · Aloi | ACC: 0.796 -> 0.916 | Helped |
| add consistency distillation (ablation) | MGCD server · Aloi | ACC: 0.859 -> 0.916 | Helped |
Surprising findings
- The diversity-measurement term Ldd contributed more than the consistency term Lcl in ablation, per the authors' analysis.
- Removing the dual autoencoder in favor of a traditional autoencoder lowers Aloi ACC from 0.916 to 0.796.
Limitations
- Privacy claim relies on graph uploads resisting model inversion, without a formal attack evaluation shown.
- Baselines reimplemented from source code with suggested parameters; results on some datasets (e.g., Animal) show very low absolute scores for baselines.
Experimental setup
Datasets: Mfeat Scene Aloi Animal Cifar10 NoisyMNIST
Benchmarks: Mfeat Scene Aloi Animal Cifar10 NoisyMNIST
Metrics: ACC NMI ARI Purity
Baselines: LMVSC SiMVC CoMVC MFLVC GCFAgg FedMVL FedDMVC FCUIF
Builds on: federated multi-view learning (FedMVL) FedDMVC MFLVC self-representation subspace clustering
Closest papers by meaning
- Bridging Gaps: Federated Multi-View Clustering in Heterogeneous Hybrid Views (NeurIPS 2024 · similarity 0.91)
- Federated Incomplete Multi-view Clustering with Globally Fused Graph Guidance (ICML 2025 · similarity 0.90)
- Scalable Federated One-Step Multi-View Clustering with Tensorized Regularization (AAAI 2025 · similarity 0.90)
- Efficient Federated Incomplete Multi-View Clustering (ICML 2025 · similarity 0.90)
- An Effective and Secure Federated Multi-View Clustering Method with Information-Theoretic Perspective (ICML 2025 · similarity 0.89)
- Multi-View Clustering by Inter-cluster Connectivity Guided Reward (ICML 2024 · 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.