Scalable spatial–geometric–temporal graph learning for radio resource management in wireless communication

Abstract

Next-generation wireless networks, such as 6G and device to device (D2D) communications, must accommodate an ever growing number of users, channels, and quality of service (QoS) requirements. This increasing complexity makes traditional optimization based approaches inadequate for real time, large-scale network operation. This thesis, presented in the form of research articles, investigates the use of graph neural networks (GNNs) for the joint tasks of power control and spectrum allocation in wireless networks. The following three main contributions are proposed: 1. A GNN CNN model capable of processing non-Euclidean interference graphs to enable efficient, scalable joint allocation, while maintaining robustness under imperfect channel state information (CSI); 2. An unsupervised spatial and geometric learning framework combining a variational spatial autoencoder and a variational graph autoencoder, which enables generalization to variable size networks without retraining; 3. An event-based temporal graph neural network that models continuous time dynamic graphs (CTDGs), allowing adaptation to high mobility environments and rapidly changing network topologies. Experimental results show that the proposed methods outperform both heuristic and conventional deep learning baselines in terms of average throughput, QoS satisfaction, and generalization capability, while also reducing computational complexity.

Date	15 May 2026
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Georges Kaddoum (Supervisor)