Graph Neural Network Empowered Wireless Communications: Fundamentals, State-of-the-Art, Challenges, and Opportunities

Deep learning (DL) has been widely applied in wireless communications to address diverse challenges, such as beam management, positioning, and channel state information (CSI) feedback. However, the traditional DL methods, for example, fully connected neural networks, convolution neural networks (CNNs), recurrent neural networks (RNNs), and so on, are constrained by limited extensibility and centralized deployment, making them unsuitable for dynamic and distributed wireless communication systems. To overcome these limitations, the emerging graph neural network (GNN) paradigm has attracted increasing attention in wireless communications. In this article, we provide a comprehensive overview of GNNs in wireless communications. Specifically, we commence by delineating the fundamentals of GNNs and the distinctions between GNNs and other traditional DL methods. Subsequently, we highlight the intrinsic properties of GNNs that render them well-suited for wireless communications, including permutation equivariance, the ability to unfold iterative algorithms, distributed deployment, and robust extensibility. Next, we delve into recent studies on GNN-empowered wireless communications by differentiating them into three main categories: node-level tasks, edge-level tasks, and graph-level tasks. Finally, the challenges and opportunities inherent in leveraging GNNs for wireless communications are identified, and future directions are discussed.