Understanding the Statistical Accuracy-Communication Trade-off in Personalized Federated Learning with Minimax Guarantees
Authors: Xin Yu, Zelin He, Ying Sun, Lingzhou Xue, Runze Li
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The theoretical result is validated on both synthetic and real-world datasets and its generalizability is verified in a non-convex setting. |
| Researcher Affiliation | Academia | 1 Department of Statistics, The Pennsylvania State University, University Park, PA 16802, USA 2 School of Electrical Engineering and Computer Science, The Pennsylvania State University, University Park, PA 16802, USA. Correspondence to: Ying Sun <EMAIL>, Lingzhou Xue <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 Federated Gradient Descent with K-Step Local Optimization Algorithm 2 Fed CLUP: Federated Learning with Constant Local Update Personalization |
| Open Source Code | Yes | Details about the experimental setup are available in Section 6.1 and Appendix D, and the complete anonymized codebase is accessible at https://github.com/ZLHe0/fedclup. |
| Open Datasets | Yes | We use the MNIST, EMNIST, CIFAR10, Sent140 an Celeb A datasets for real data analysis. |
| Dataset Splits | No | The paper mentions distributing data across clients and using training and testing sets, but does not provide specific details on the percentage splits or sample counts for training, validation, and test sets. For example, Section D.1 states: "For the real datasets, we report training loss and testing accuracy." and "To impose statistical heterogeneity, we distribute the data across clients in a way that each client only has access to a fixed number of classes." These statements indicate the use of splits but lack the concrete, specific information required for reproducibility. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU types, or memory amounts used for running the experiments. |
| Software Dependencies | No | The paper mentions software like "Py Torch" for implementation but does not specify any version numbers for PyTorch or other libraries. For instance, in Section D.1: "For the MNIST dataset, all models are implemented in Py Torch, and optimization is performed using the stochastic gradient descent (SGD) solver." |
| Experiment Setup | Yes | For the synthetic dataset, aligned with Theorem 2, we set the global step size γ = (λ+L)/(λL) and the local step size η = (L+λ) 1. For the real dataset, the global learning rate was set to 1/λ, while the local learning rate was set to 0.01, chosen via grid search. For the MNIST dataset, we implemented logistic regression using the SGD solver with 5 epochs, a batch size of 32, and 20 total runs. |