Research map / Federated learning and privacy / Personalized and heterogeneous FL
Federated Incomplete Multi-view Clustering with Globally Fused Graph Guidance
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 FIMCFG, a federated incomplete multi-view clustering method that uses a server-fused graph to guide a dual-head GCN encoder and fusion module on each client, with pseudo-label clustering and silhouette-weighted aggregation.
Problem
Federated multi-view clustering typically uses global information only for pseudo-label supervision of clustering, not for upstream feature extraction, and handles missing views poorly (incomplete data is less explored in federated MVC).
Approach
A dual-head GCN encoder per client extracts global-graph and local-graph underlying features; a fusion module maps them to high-level features under the guidance of a globally fused graph built at the server. Global graph structure migration repairs local graphs for missing samples, and silhouette scores weight server-side graph and feature fusion.
Key results
- At missing rate 0.5, FIMCFG reaches ACC 0.952 on HW, 0.444 on Scene-15, 0.293 on Landuse-21, and 0.802 on 100Leaves, the best in each column of the reported table.
- Ablation on Scene-15 (delta=0.5) shows removing all fused-graph components drops ACC from 0.444 to 0.247.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FIMCFG instead of the strongest baseline, FCUIF | Federated incomplete multi-view clustering · HW (δ=0.5) | ACC: 0.917 -> 0.952 | Helped |
| Use FIMCFG instead of the strongest baseline, FCUIF | Federated incomplete multi-view clustering · Scene-15 (δ=0.5) | ACC: 0.410 -> 0.444 | Helped |
| Use FIMCFG instead of the strongest baseline, FCUIF | Federated incomplete multi-view clustering · Landuse-21 (δ=0.5) | ACC: 0.232 -> 0.293 | Helped |
| Use FIMCFG instead of the strongest baseline, GIMVC | Federated incomplete multi-view clustering · 100Leaves (δ=0.5) | ACC: 0.688 -> 0.802 | Helped |
| Add global graph structure migration (ablation) | FIMCFG · Scene-15 (δ=0.5) | ACC: 0.357 -> 0.444 | Helped |
| Add the feature fusion module (ablation) | FIMCFG · Scene-15 (δ=0.5) | ACC: 0.431 -> 0.444 | Helped |
| Add global graph guidance (ablation) | FIMCFG · Scene-15 (δ=0.5) | ACC: 0.412 -> 0.444 | Helped |
| Add the global graph encoder and related fused-graph components (ablation) | FIMCFG · Scene-15 (δ=0.5) | ACC: 0.247 -> 0.444 | Helped |
| Add global graph structure migration (ablation) | FIMCFG · Landuse-21 (δ=0.5) | ACC: 0.213 -> 0.293 | Helped |
| Add the feature fusion module (ablation) | FIMCFG · 100Leaves (δ=0.5) | ACC: 0.758 -> 0.802 | Helped |
| Add global graph guidance (ablation) | FIMCFG · HW (δ=0.5) | ACC: 0.912 -> 0.952 | Helped |
| Add the global graph encoder and related fused-graph components (ablation) | FIMCFG · 100Leaves (δ=0.5) | ACC: 0.430 -> 0.802 | Helped |
Surprising findings
- Removing global graph structure migration (component A) lowers ACC on Scene-15 from 0.444 to 0.357 while NMI stays near 0.407, suggesting the components affect metrics unevenly.
Limitations
- Absolute clustering accuracy on Landuse-21 and Scene-15 remains low (around 0.3 to 0.45 at 50% missing).
- Experiments use fixed-size synthetic missing patterns and a single heterogeneity construction rather than real federated deployments.
Experimental setup
Datasets: Scene-15 HandWritten Landuse-21 100Leaves
Benchmarks: HandWritten Scene-15 Landuse-21 100Leaves
Metrics: ACC NMI ARI
Baselines: FCUIF FedMVFPC HFMVC DIMVC DSIMVC GIGA GIMVC CDIMC-net MRL_CAL
Builds on: Completer GIMVC FCUIF
Closest papers by meaning
- Efficient Federated Incomplete Multi-View Clustering (ICML 2025 · similarity 0.91)
- Graph Consistency and Diversity Measurement for Federated Multi-View Clustering (AAAI 2025 · similarity 0.90)
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.89)
- Bridging Gaps: Federated Multi-View Clustering in Heterogeneous Hybrid Views (NeurIPS 2024 · similarity 0.89)
- Incomplete and Unpaired Multi-View Graph Clustering with Cross-View Feature Fusion (AAAI 2025 · similarity 0.89)
- Towards Resource-friendly, Extensible and Stable Incomplete Multi-view Clustering (ICML 2024 · similarity 0.88)
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.