GNN-Enabled Max-Min Fair Beamforming

This paper investigates a graph neural network (GNN)-enabled beamforming design to achieve max-min fairness for multi-user multiple-input-single-output (MU-MISO) networks. By modelling the MU-MISO network as a directed graph with defined node and edge features, the max-min rate problem is transformed into a graph optimization problem. We then solve the problem by a new GNN-based model named complex edge graph attention networks (CEGAT). The proposed CEGAT directly learns the mapping between channel state information and beamforming vectors. With a power adjustment unit to address the power budget constraint, CEGAT can be trained in an unsupervised manner. Numerical results validate the proposed CEGAT in terms of optimality, scalability to number of users and power budget and inference time.