1: \begin{abstract}

2: Federated learning (FL) is a machine learning paradigm that

3: targets model training without gathering the local data dispersed

4: over various data sources. Standard FL, which employs a single

5: server, can only support a limited number of users, leading

6: to degraded learning capability. In this work, we consider a

7: multi-server FL framework, referred to as \emph{Confederated

8: Learning} (CFL), in order to accommodate a larger number of

9: users. A CFL system is composed of multiple networked edge

10: servers, with each server connected to an individual set of users.

11: Decentralized collaboration among servers is leveraged to harness

12: all users' data for model training. Due to the potentially massive

13: number of users involved, it is crucial to reduce the communication

14: overhead of the CFL system. We propose a stochastic

15: gradient method for distributed learning in the CFL framework.

16: The proposed method incorporates a conditionally-triggered user

17: selection (CTUS) mechanism as the central component to effectively

18: reduce communication overhead. Relying on a delicately designed

19: triggering condition, the CTUS mechanism allows each server to

20: select only a small number of users to upload their gradients,

21: without significantly jeopardizing the convergence performance of

22: the algorithm. Our theoretical analysis reveals that the proposed

23: algorithm enjoys a linear convergence rate. Simulation results

24: show that it achieves substantial improvement over state-of-the-art

25: algorithms in terms of communication efficiency.

26: \end{abstract}