STAGNN: a spatial-temporal attention graph neural network for network traffic prediction

Accurate and real-time traffic prediction can reasonably allocate the resources of communication networks and effectively improve the communication quality of networks. However, the complex topology and highly dynamic nature of communication networks pose new challenges for traffic prediction. To be able to effectively obtain the temporal correlation and spatial dependency of network traffic and mask the redundant traffic features, we propose a spatial-temporal attention graph neural network (STAGNN). The STAGNN combines the graph attention network (GAT) and the time series model informer, where GAT is used to learn the complex spatial dependencies of network topology and informer is used to learn the dynamic temporal correlation of network traffic. Also in learning, we introduce the multi-headed attention mechanism enabling STAGNN to quickly select high-value network traffic information using limited attention resources. The experimental results demonstrate that STAGNN has better prediction performance compared with other existing methods.