A Network Traffic Prediction Model Based on Layered Training Graph Convolutional Network

Routing deployment and resource scheduling in communication networks require accurate traffic prediction. Neural network-based models that extract the time-correlated or space-correlated features of traffic flow have been developed for traffic prediction. The conventional model that extracts space-correlated features of traffic flow have the problem of high computational complexity and long training time which limits the model's application on rapid routing deployment. This paper therefore proposes a layered training graph convolutional network (LT-GCN) to decrease the training time greatly with the nearly same prediction accuracy as graph convolutional network (GCN). Instead of training on parameters in all hidden layers simultaneously, LT-GCN develops a new layer-by-layer training pattern for multiple hidden layers to degrade the computational complexity in training process. LT-GCN is then further integrated with gated recurrent unit (GRU) that is called LTGG model to achieve the joint extraction of time-correlated and space-correlated features of traffic flow for more accurate prediction. Experimental results demonstrate that LT-GCN outperforms the classical GCN model on training time and LTGG exhibits greater performance than other benchmark models on prediction accuracy.