Abstract: Aiming at the problem that the hybrid beamforming in millimeter wave (mm-Wave) massive multiple input multiple output (MIMO) system leads to excessive communication overhead in centralized machine learning (CML), a hybrid beamforming and resource allocation optimization method based on contract theory under hierarchical federated learning (FL) was proposed. First, the system overhead was calculated for multiple users in a hierarchical system, and the maximum benefit of the system was realized by optimizing the allocation of resources. Then, the allocated resources were used to train the channel data and corresponding precoder data in the back propagation neural network (BPNN). The model was aggregated at the edge server (ES) by collecting weights and parameters from users and uploaded to the cloud server (CS) to aggregate after reaching a certain accuracy until obtaining an optimal model. The simulation results show that the resource optimization greatly reduces the communication overhead, and hybrid beamforming based on FL not only achieves almost the same sum rate as to CML, but also outperforms orthogonal matching pursuit (OMP) hybrid beamforming and fully digital beamforming.