Edge Feature Empowered Graph Attention Network for Sum Rate Maximization in Heterogeneous D2D Communication System

Qiushi Sun, Yang He, Yin Li, Ovanes L. Petrosian

Abstract: Wireless mobile technology has advanced to a stage where data rates have significantly increased, and the capacity for device connectivity is poised for exponential growth. Achieving high data rate transmission in complex communication systems has become critical. Traditional iterative methods face challenges due to their high computational complexity and time-consuming processes. Graph neural networks (GNNs) have shown great potential in optimizing network resource allocation problems. GNNs exhibit strong learning capabilities in handling graph-structured data and effectively utilizing the topology of communication networks. However, current approaches primarily focus on node features, often neglecting the valuable information in edge features. We propose the Edge Feature Empowered Jumping Knowledge Graph Attention Network (EJGAT) architecture to break this limitation. This model employs a hierarchical attention mechanism and jumping knowledge connection to enhance learning capability and mitigate the over-smoothing problem. Simulation experiments illustrate that the proposed algorithm outperforms benchmarks in terms of both average sum rate and learning efficiency. Comprehensive experimental validation further highlights the algorithm’s potential applicability to various system parameters and its robustness to corrupted input features.