Beamforming and Load-Balanced User Association in RIS-Aided mmWave Systems via Adaptive Attention Graph Neural Networks

This paper investigates the joint optimization of user association, base station (BS) beamforming, and reconfigurable intelligent surface (RIS) phase adjustment in an RIS-aided multi-BS multi-user-equipment (UE) millimeter wave (mmWave) network with load-balancing considerations. To address this complex problem, we propose a graph neural network (GNN)-based approach with enhanced generalizability to both varying numbers of BSs and UEs. An adaptive attention-based aggregation (AAA) mechanism is incorporated to mitigate oversmoothing in deep GNNs. By using uplink pilots as input, the proposed method avoids the need for explicit channel state information (CSI). Simulation results demonstrate the proposed scheme’s superior sum-rate performance compared to both optimization-based and learning-based benchmarks while satisfying load-balancing requirements, quantify the effectiveness of AAA in addressing oversmoothing, demonstrate the proposed scheme’s robustness to pilot correlation, and highlight the load-balancing benefits of deploying RISs.