Interaction-Aware Gaussian Weighting for Clustered Federated Learning
Authors: Alessandro Licciardi, Davide Leo, Eros Fanı̀, Barbara Caputo, Marco Ciccone
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments on benchmark datasets show that Fed GWC outperforms existing FL algorithms in cluster quality and classification accuracy, validating the efficacy of our approach. Through extensive experiments on both benchmark (Caldas et al., 2018) and large-scale datasets (Hsu et al., 2020), we demonstrate that Fed GWC outperforms existing clustered FL algorithms in terms of both accuracy and clustering quality. In this section, we present the experimental results on widely used FL benchmark datasets (Caldas et al., 2018) including real-world datasets (Hsu et al., 2020), comparing the performance of Fed GWC with other baselines from the literature, including standard FL algorithms and clustering methods. |
| Researcher Affiliation | Academia | 1Department of Mathematical Sciences, Polytechnic University of Turin, Italy 2Istituto Nazionale di Fisica Nucleare (INFN), Sezione di Torino, Turin, Italy 3Department of Computing and Control Engineering, Polytechnic University of Turin, Italy 4Basque Center for Applied Mathematics (BCAM), Bilbao, Spain 5Vector Institute, Toronto, Ontario, Canada. |
| Pseudocode | Yes | Algorithm 1 Fed GWC 1: Input: K, T, S, αt, ϵ, |Pt|, K 2: Output: C(1), . . . , C(Ncl) and θ(1), . . . , θ(Ncl) 3: Initialize N 0 cl 1 ... Algorithm 2 Fed GW Cluster 1: Input: P, nmax, K( , ) 2: Output: cluster labels yncl, and number of clusters ncl 3: Extract UPVs vj k, vk j from P for any k, j |
| Open Source Code | Yes | Code is available at https: //github.com/davedleo/Fed GWC |
| Open Datasets | Yes | Our experiments on benchmark datasets show that Fed GWC outperforms existing FL algorithms in cluster quality and classification accuracy, validating the efficacy of our approach. Through extensive experiments on both benchmark (Caldas et al., 2018) and large-scale datasets (Hsu et al., 2020), we demonstrate that Fed GWC outperforms existing clustered FL algorithms in terms of both accuracy and clustering quality. Furthermore, Fed GWC can be integrated with any robust FL aggregation algorithm to provide additional resilience against data heterogeneity. |
| Dataset Splits | Yes | Each client has its own local train and test sets. We conduct experiments on Cifar100 (Krizhevsky et al., 2009). As a comparison, we also run experiments on the simpler Cifar10 dataset (Krizhevsky et al., 2009). Cifar10 and Cifar100 are distributed among K clients using a Dirichlet distribution (by default, we use α = 0.05 for Cifar10 and α = 0.5 for Cifar100) to create highly imbalanced and heterogeneous settings. By default, we use K = 100 clients with 500 training and 100 test images. The classification model is a CNN with two convolutional blocks and three dense layers. Additionally, we perform experiments on the Femnist dataset (Le Cun, 1998), partitioned among 400 clients using a Dirichlet distribution with α = 0.01. |
| Hardware Specification | No | The paper mentions 'high-performance computing resources' in the acknowledgements but does not specify any particular hardware components such as CPU/GPU models or memory details used for the experiments. |
| Software Dependencies | No | The paper does not explicitly list any software dependencies with specific version numbers. It mentions using SGD and Le Net5, but no versions are provided. |
| Experiment Setup | Yes | Local training on each client uses SGD with a learning rate of 0.01, weight decay of 4e-4, and batch size 64. The number of local epochs is 1, resulting in 7 batch iterations for Cifar10 and Cifar100 and 8 batch iterations for Femnist. The number of communication rounds is set to 3,000 for Femnist, 10,000 for Cifar10 and 20,000 for Cifar100, with a 10% client participation rate per cluster. For Fed GWC we tuned the hyper-parameter β {0.1, 0.5, 1, 2, 4}, i.e. the spread of the RBF kernel, and we set the tolerance ϵ to 10-5, constant value αt = α equal to the participation rate, i.e. 10%. |