Research map / Federated learning and privacy / Personalized and heterogeneous FL
Interaction-Aware Gaussian Weighting for Clustered Federated 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 FedGWC, a clustered federated learning method that clusters clients via Gaussian-reward estimates of their empirical loss trajectories and spectral clustering on an interaction-derived affinity matrix, plus a class-imbalance-adjusted Wasserstein clustering metric (WAS/WADB).
Problem
Federated learning under data heterogeneity and class imbalance, where a single global model degrades; clustered FL needs to group clients with similar distributions without heavy communication overhead or sensitive hyperparameters, and existing cluster metrics ignore class imbalance.
Approach
Transform each client's empirical loss trajectory into Gaussian rewards relative to the federation mean, running-average them into an interaction matrix, convert to an affinity via RBF over unbiased perception vectors, and spectrally cluster with Davies-Bouldin-based selection of cluster count; the Wasserstein Adjusted Score adjusts standard clustering metrics by comparing sorted class-frequency v
Key results
- On CIFAR-100 (Dirichlet alpha=0.5, 100 clients), FedGWC reaches 53.4% accuracy with WAS 0.1 and WADB 2.4, matching FeSem's accuracy and beating IFCA (47.5%) and FedAvg (41.6%).
- On Google Landmarks, FedGWC with FedAvg aggregation reaches 57.4% accuracy versus 49.4% for IFCA and 40.5% for FedAvg.
- On iNaturalist, FedGWC reaches 47.8% accuracy versus 45.8% for IFCA and 45.3% for FedAvg.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedGWC client clustering with FedAvg aggregation instead of standard FedAvg | FedAvg · Cifar100 | accuracy: 41.6 -> 53.4 | Helped |
| Use FedGWC client clustering with FedAvg aggregation instead of standard FedAvg | FedAvg · Femnist | accuracy: 76.6 -> 76.1 | Hurt |
| Use FedGWC clustering instead of FeSem | Clustered FL with FedAvg aggregation · Cifar100 | WAS: -0.3 -> 0.1 | Helped |
| Use FedGWC clustering instead of IFCA | Clustered FL with FedAvg aggregation · Cifar100 | WADB: 5.2 -> 2.4 | Helped |
| Use FedGWC clustering instead of standard FedAvg | FedAvg · Google Landmarks | accuracy: 40.5 -> 57.4 | Helped |
| Use FedGWC clustering instead of standard FedAvg | FedAvg · iNaturalist | accuracy: 45.3 -> 47.8 | Helped |
| Use FedGWC clustering instead of IFCA | Client clustering · Cifar100 (10, 0, 90) class heterogeneity | WAS: -0.9 -> 0.1 | Helped |
| Use FedGWC clustering instead of FeSem | Client clustering · Cifar100 (10, 90, 0) class heterogeneity | WAS: 0.2 -> 0.4 | Helped |
| Use FedGWC clustering instead of IFCA | Client clustering · Cifar100 (40, 30, 30) class heterogeneity | WADB: 1.0 -> 0.9 | Helped |
| Use FedGWC clustering instead of FeSem | Client clustering · Cifar100 (Clean, Noise) visual domains | Rand Index: 0.49 -> 1.0 | Helped |
| Use FedGWC clustering instead of FeSem | Client clustering · Cifar100 (Clean, Blur) visual domains | Rand Index: 0.51 -> 1.0 | Helped |
| Use FedGWC clustering instead of FeSem | Client clustering · Cifar100 (Clean, Noise, Blur) visual domains | Rand Index: 0.55 -> 0.6 | Helped |
Surprising findings
- On FEMNIST, FedAvgM (83.3%) outperforms all clustered methods, including FedGWC (76.1%), and FedGWC has a negative WAS (-0.2), so clustering did not help there.
- On CIFAR-100 the Rand Index for FedGWC reaches 1.0 on the two-domain settings, while IFCA with one cluster scores 0.5.
Limitations
- On FEMNIST, clustered methods underperform standard FedAvg and FedAvgM, and the paper concedes clustering is not always beneficial.
- Clustering quality depends on the RBF kernel hyperparameter beta and the threshold epsilon, which the paper says requires sensitivity analysis.
- Experiments use small CNNs and simulated Dirichlet partitions; the proposed convergence results rely on stationarity assumptions and are established for individual weights rather than the full training dynamics.
Experimental setup
Datasets: CIFAR-100 CIFAR-10 FEMNIST Google Landmarks iNaturalist
Benchmarks: CIFAR-100 FEMNIST Google Landmarks iNaturalist
Metrics: balanced accuracy Wasserstein Adjusted Silhouette Score (WAS) Wasserstein Adjusted Davies-Bouldin Score (WADB) Rand Index
Baselines: IFCA FeSEM CFL FedAvg FedAvgM FedProx FairAvg
Builds on: IFCA Clustered FL FedAvg Robbins-Monro stochastic approximation
Closest papers by meaning
- Enhancing Clustered Federated Learning: Integration of Strategies and Improved Methodologies (ICLR 2025 · similarity 0.91)
- Clustered Federated Learning via Gradient-based Partitioning (ICML 2024 · similarity 0.90)
- FLUX: Efficient Descriptor-Driven Clustered Federated Learning under Arbitrary Distribution Shifts (NeurIPS 2025 · similarity 0.89)
- FedAWA: Adaptive Optimization of Aggregation Weights in Federated Learning Using Client Vectors (CVPR 2025 · similarity 0.88)
- A Bayesian Approach for Personalized Federated Learning in Heterogeneous Settings (NeurIPS 2024 · similarity 0.88)
- FedLMT: Tackling System Heterogeneity of Federated Learning via Low-Rank Model Training with Theoretical Guarantees (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.