Heterogeneous Graph Neural Network for Beamforming Design in Cell-Free Massive MIMO with Underlaid D2D Maritime Systems

Considering the critical issue of interference management in cell-free massive MIMO with underlaid device-todevice (CF-mMIMO-D2D) maritime systems, in this paper, we propose a novel approach, namely CFD-HetGNN, utilizing the heterogeneous graph neural network (HetGNN) to deal with the joint optimization of beamforming and power allocation for both the CF-mMIMO users and D2D pairs. By modeling the CF-mMIMO-D2D as a heterogeneous graph, we take advantage of the HetGNN to capture the complex interactions among access points, mobile stations and D2D pairs, enabling an efficient solution of interference mitigation. Simulation results demonstrate that the proposed CFD-HetGNN significantly improves the spectral efficiency and achieves superior interference control compared to the traditional approach of matched filtering with equal power allocation (MF-EPA) as well as the learning-assisted methods of convolutional neural network (CNN), deep neural network (DNN) and long short-term memory (LSTM). Additionally, the CFD-HetGNN possess the robust scalability, making it applicable to the large-scale networks with low complexity.