Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Rethinking Fair Federated Learning from Parameter and Client View

Authors: Kaiqi Guan, Wenke Huang, Xianda Guo, Yueyang Yuan, Bin Yang, Mang Ye

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments on single-domain and cross-domain scenarios. With ablations, we validate the efficacy of Fed PW and the indispensability of modules.
Researcher Affiliation Academia 1 National Engineering Research Center for Multimedia Software, Institute of Artificial Intelligence, Hubei Key Laboratory of Multimedia and Network Communication Engineering, School of Computer Science, Wuhan University, Wuhan, China. EMAIL
Pseudocode Yes Algorithm 1: Fed PW
Open Source Code Yes The code is available at https://github.com/guankaiqi/Fed PW.
Open Datasets Yes Following [18, 21, 43], we evaluate our method on single-domain datasets Fashion-Mnist [54], Cifar10 [28], Cifar100, and cross-domain datasets Digits [29] and Office-Caltech [11].
Dataset Splits Yes To simulate heterogeneous clients in FL, we consider three scenarios: (1) Dir(α): We simulate m clients in Dirichlet heterogeneous partition. The smaller α is, the more imbalanced the local distribution is. (2) Pat-1: It constructs a difficult data-island scenario where each client only has data from one class. (3) Pat-2: We follow Fed Avg to build pathological non-IID data where each client has data from two classes. For the single-domain setting, we employ 100 clients for 3,000 communication epochs, where all federated learning methods exhibit minimal or no accuracy improvement beyond this point. Each epoch involves 10% client participation. We use the SGD optimizer with a learning rate of 0.1 and a batch size of 50. For the cross-domain setting, we allocate 20 clients per task and equal clients per domain, with clients randomly assigned to domains.
Hardware Specification Yes We conduct experiments on the NVIDIA 3090Ti.
Software Dependencies No The paper mentions using 'SGD optimizer' but does not specify any software libraries or frameworks with version numbers for implementation.
Experiment Setup Yes We use the SGD optimizer with a learning rate of 0.1 and a batch size of 50. For the cross-domain setting, we allocate 20 clients per task and equal clients per domain, with clients randomly assigned to domains. The training runs for E = 200 communication epochs with T = 10 local updates per round. Each epoch involves all clients. The SGD uses a learning rate of 0.001, and momentum is 0.9. The batch sizes are 64 for Digits and 16 for Office-Caltech.