Communication efficiency arises as a necessity in federated learning due to limited communication bandwidth. To this end, the present paper develops an algorithmic framework where an ensemble of pre-trained models is learned. At each learning round, the server selects a subset of pre-trained models to construct the ensemble model based on the structure of a graph, which characterizes the server's confidence in the models. Then only the selected models are transmitted to the clients, such that certain budget constraints are not violated. Upon receiving updates from the clients, the server refines the structure of the graph accordingly. The proposed algorithm is proved to enjoy sub-linear regret bound. Experiments on real datasets demonstrate the effectiveness of our novel approach.
View on arXiv