Abstract:To address the challenges of communication overhead between clients and server, as well as the difficulties in achieving convergence and maintaining high accuracy in the central model during federated learning, this study proposes a federated communication framework named FedHNC, based on a node selection strategy. Introducing a probabilistic selection method on the server, the framework identifies and differentiates nodes and assigns them various probability attributes to dynamically select higher-quality nodes for communication. Additionally, it incorporates a personalized communication process where clients compress local updates by using a Top-K compressor with error feedback. In the aggregation phase, the traditional weighted averaging is replaced by an element-based aggregation method.The momentum and error correction are also combined to optimize the global update. The compressed global parameters are then compressed and distributed to clients for the next training round. Experiments are conducted on two datasets and three models. The results demonstrate that while maintaining center model accuracy, FedHNC reduces communication overhead by 41% on average and improves convergence speed by approximately 73% compared to FedAvg. This framework addresses the communication burden overlooked by baseline models and shows significant application potential in improving federated communication efficiency, accelerating model convergence, and ensuring model effectiveness.