Graph Neural Network-Based Unified Beamforming and User Selection for MU-MISO

Classical beamforming algorithms face challenges such as increased complexity with more users and base station antennas (BSAs), as well as the requirement of accurate channel state information (CSI). When the number of users exceeds the number of BSAs, user selection becomes necessary, typically managed in zero-forcing beamforming using semi-orthogonal user selection. However, integrating this approach with deep learning-based beamforming remains underexplored. In this paper, we propose a deep learning-based unified beamforming and user selection method that scales efficiently with the number of users and BSAs. This approach jointly optimizes beamforming and user selection by introducing a user selection number pursuit algorithm and an attention-based message aggregation. Simulation results show that the proposed method has better sum rates than conventional methods with less computational complexity. Additionally, it demonstrates robustness against imperfect CSI compared to conventional methods.