Enhancing Clustered Federated Learning: Integration of Strategies and Improved Methodologies
Authors: Yongxin Guo, Xiaoying Tang, Tao Lin
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive numerical evaluations, we demonstrate the effectiveness of our clustering framework and the enhanced components. Extensive experiments on different datasets (CIFAR10, CIFAR100, and Tiny-Imagenet) and various architectures (Mobile Net-V2 and Res Net18) demonstrate the effectiveness of our framework and the improved components of HCFL+. |
| Researcher Affiliation | Academia | 1School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen 518172, China 2Shenzhen Institute of Artificial Intelligence and Robotics for Society (AIRS), Shenzhen, China 3Guangdong Provincial Key Laboratory of Future Networks of Intelligence, Shenzhen, China 4School of Engineering, Westlake University 5Research Center for Industries of the Future, Westlake University |
| Pseudocode | Yes | Algorithm 1 Algorithm Framework of HCFL. Algorithm 2 Algorithm Framework of HCFL+ Algorithm 3 Local Updates of HCFL+ Algorithm 4 Cluster Adding of HCFL+ Algorithm 5 Cluster Removing of HCFL+ |
| Open Source Code | Yes | Our code is available at https://github.com/LINs-lab/HCFL. |
| Open Datasets | Yes | Extensive experiments on different datasets (CIFAR10, CIFAR100, and Tiny-Imagenet) and various architectures (Mobile Net-V2 and Res Net18) demonstrate the effectiveness of our framework and the improved components of HCFL+. For label distribution shifts, we employ LDA with α = 1.0, as introduced by (Yoshida et al., 2019; Hsu et al., 2019; Reddi et al., 2021). For feature distribution shifts, we adopt the methodology from CIFAR10-C and CIFAR100-C creation (Hendrycks & Dietterich, 2019). |
| Dataset Splits | Yes | Unless specifically mentioned, we divide the datasets into 100 clients and execute all algorithms for 200 communication rounds. We establish clients with three types of distribution shifts. For label distribution shifts, we employ LDA with α = 1.0, as introduced by (Yoshida et al., 2019; Hsu et al., 2019; Reddi et al., 2021). We split datasets to 100 clients by default. Validation Accuracy for evaluating personalization: The average accuracy on local validation datasets that match the distribution of local training sets. (2) Test Accuracy evaluating generalization: The average accuracy on global shared test datasets. |
| Hardware Specification | Yes | The experiments are conducted on single NVIDIA RTX 3090 GPUs. |
| Software Dependencies | No | The paper discusses various FL algorithms and model architectures like Mobile Net-V2 and Res Net18 but does not provide specific version numbers for software dependencies such as programming languages or libraries. |
| Experiment Setup | Yes | Unless specifically mentioned, we divide the datasets into 100 clients and execute all algorithms for 200 communication rounds. The learning rates are chosen in {0.03, 0.06, 1.0}, and we report the best results for each algorithm. We conducted all experiments using Mobile Net-V2 (Sandler et al., 2018) and results on Res Net18 defer to Table 8 of Appendix E. We set µ = 0.4, and choose ρ = {0.05, 0.1, 0.3}. |