logo
CanyamAI summaries for academic research
language
MobileMobile
  • Phone FeatureWeChat Mini Program
EnglishEnglish
noticeNotice
BlogBlog
delete
arrow
Return

LOGO-CL: Accelerating semi-supervised federated learning in edge computing

delete2025-07-04
delete0
PRE
AI
Y
Yang Xu
Q
Q. Y. Wang
H
Hongli Xu
Y
Yunming Liao
L
Liusheng Huang
X
Xin Hang
DOI：10.1016/j.comnet.2025.111496delete
deleteOriginal
deleteOriginal request for help
deleteShare
deleteSave
Abstract

Abstract

En 中文
Federated learning (FL) has gained important attention for training models across clients in edge computing. The scarcity of labeled data poses critical challenges in professional fields (e.g., medical diagnosis) for FL, motivating the development of federated semi-supervised learning (FSSL) to exploit unlabeled data. Though previous works can make full use of unlabeled data, they inevitably face two inherent challenges: limited communication resource and data heterogeneity. Existing FSSL approaches typically assign fixed values for global updating frequency and/or local updating frequency during training, and ignore the impact of unlabeled data distribution, which severely hinder convergence stability and training efficiency. To address these issues, we present a novel FSSL framework, termed LOGO-CL. Specifically, we jointly optimize both local and global updating frequencies by analyzing the coupled impact of these two hyper-parameters on the training process. We develop a multi-armed bandit (MAB) based online algorithm to adaptively determine diverse local updating frequencies as well as appropriate global updating frequency, so as to improve training efficiency. Furthermore, LOGO-CL leverages contrastive learning technique to alleviate the impact of unlabeled data heterogeneity by incorporating regularization terms into both global and local training processes, ensuring robustness, even with imbalanced data distribution. Extensive experiments demonstrate that LOGO-CL achieves the model training speedup by 2.3 × and reduces communication cost by 44% when achieving the same target accuracy, compared with baselines. Moreover, LOGO-CL improves model accuracy up to 4.8% on the datasets with moderate noise.

Journal

Computer Networks cover
Computer Networks
IF:
4.6
Papers:
1.5K
Citations:
1.6W

Organization

No organization information available