Decentralized Training of Graph Neural Networks in Mobile Systems for Power Control

Graph neural networks (GNNs) have been used for optimizing resource allocation due to their potential in scalability and size generalizability. To facilitate their application to large- scale wireless systems, decentralized inference with GNNs has been investigated recently. Yet decentralized training of GNNs at wireless nodes, which can alleviate the computing load at central server and protect privacy of users, has never been studied. In this paper, we strive to train GNNs in mobile systems in a decentralized manner, by taking power control optimization for interference coordination as an example. We present a framework for decentralized training of GNNs at wireless nodes, and propose two algorithms to tackle the challenge of training GNNs over dynamic graph topology. Simulation results show that the power control policy learned by the GNN performs very close to decentralized numerical algorithms with lower signaling overhead for inference.