Edge-Wise Gated Graph Neural Network for User Association in Massive URLLC

Abstract

To support ultra-reliable and low-latency communications with massive connections, we develop a novel user association algorithm in mobile edge computing with grant-free random access based on an edge-wise gated graph neural network (EG-GNN). The parameters of the EG-GNN are trained in an unsupervised manner by minimizing the overall packet loss probability, including the decoding error probability of short packet transmissions, the collision probability of random access, and the processing delay violation probability in edge servers. By representing the wireless network by a bipartite graph with base station nodes and device nodes, we apply the EG-GNN to user association, where the "gate" of an edge is defined as the probability that a device will associate with a BS. The values gates are determined by a fully connected neural network at each device node. To improve the training efficiency, analytical results in wireless communications and queueing theory are exploited to update node features. Simulation and analytical results demonstrate that the proposed EG-GNN approach outperforms existing benchmarks in terms of overall reliability with linear complexity.