Optimizing Model Dissemination for Hierarchical Clustering Learning in Edge Computing

Hierarchical clustering learning (HCL) extends traditional parameter server-based distributed learning by clustering heterogeneous user equipments (UEs) via cluster nodes (CNs) located at the edge of the network. Currently, most vanilla model dissemination strategies in distributed learning rely on one-to-many transmissions, inevitably consuming excessive precious bandwidth resources. Consequently, communication-efficiency becomes crucial for HCL in resource-constrained edge networks. In this paper, we propose a multistage cooperative model dissemination strategy to sequentially determine the subsets of CNs that can concurrently transmit models during individual scheduling stages, thereby improving communication efficiency in HCL. We formulate the strategy design as an optimization problem to minimize the maximum completion time of the slowest straggler in communication rounds, while accurately clustering UEs to CNs with similar data distributions. To make sequential and combinatorial decisions in individual stages, we develop an online learning algorithm, called sequential combinatorial multi-armed bandit (SCMAB). The SCMAB enables learning a multistage cooperative model dissemination strategy via an asymptotically optimal approach. Furthermore, the SCMAB dynamically re-clusters UEs to appropriate CNs, according to the similarity of UEs' data distribution. The simulation results indicate that compared to traditional transmission strategies, the proposed strategy improves communication efficiency by 2.11% to 5.57%, while achieving comparable and even higher learning accuracy.