SDGNet: A Handover-Aware Spatiotemporal Graph Neural Network for Mobile Traffic Forecasting
Yini Fang, Salih Ergüt, Paul Patras
Abstract
Accurate mobile traffic prediction at city-scale is becoming increasingly important as data demand surges and network deployments become denser. How mobile networks and user mobility are modelled is key to high-quality forecasts. Prior work builds on distance-based Euclidean (grids) or invariant graph representations, which cannot capture dynamic spatiotemporal correlations with high fidelity. In this letter we propose SDGNet, a handover-aware spatiotemporal graph neural network that hinges on Dynamic Graph Convolution and Gated Linear Units to predict traffic consumption over short, medium and long time-frames. Experiments with a real-world dataset demonstrate SDGNet outperforms state-of-the-art neural model, attaining up to <inline-formula> <tex-math notation="LaTeX">$4\times $ </tex-math></inline-formula> lower prediction errors.