Structure-Driven Graph Neural Networks for Scalable Multi-Group Multicasting

We consider downlink multicast beamforming in massive MIMO systems under imperfect channel state information (CSI). Existing optimization-based methods typically assume perfect CSI and can impose a significant computational burden in real-time applications. In this paper, we propose a robust and efficient learning-based approach to generate beamformers effectively. We obtain the optimal beamforming structure under imperfect CSI and propose a graph neural network (GNN) driven by this structure. The GNN model effectively captures the interactions among devices within the multicast topology while focusing the learning process within a lower-dimensional subspace, leading to enhanced overall performance. Simulation results show that the model can efficiently learn to solve the multicasting problem and is generalizable to different numbers of groups and users without requiring retraining. It outperforms existing learning-based models and closely matches the performance of the optimization-based method, but with significantly reduced computation time.