Lcsa-fed: a low cost semi-asynchronous federated learning based on lag tolerance for services QoS prediction

As a distributed training method, federated learning (FL) has been widely used in the field of quality-of-service (QoS) prediction. However, existing FL-based QoS prediction methods ignore the unreliability of end devices, which leads to wasted training resources and high communication costs. Considering the instability of end devices in real training environments, we propose a low-cost semi-asynchronous federated learning method (LCSA-Fed) based on lag tolerance to overcome the lower convergence rate and suboptimal prediction accuracy of models. LCSA-Fed is able to reduce model communication costs and training costs by tolerating relatively lagging local models. At the same time, we employ innovations in both the user selection phase and the model aggregation phase to improve prediction accuracy while reducing overhead. By conducting relevant validation experiments on a publicly available QoS dataset, we conclude that our model LCSA-Fed can reduce overhead by about 50% and improve prediction accuracy by 15.86%similar to\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}32.05%.