Spatio-temporal communication network traffic prediction method based on graph neural network

The function of network traffic prediction plays an important role in many network operations such as security, path planning and congestion control etc. Most traditional traffic prediction methods only consider temporal correlation but ignore spatial correlation, which may result in limited accuracy. In this paper, we propose an effective traffic prediction method based on the graph multi-head attention convolution neural network model, termed as FlowDiviner, which combines graph convolutional network (GCN) and multi-head attention mechanism in its encoder-decoder architecture. Specifically, GCN is used to extract spatial correlation from complex network topologies and multi-head attention mechanism is used to capture dynamic temporal correlations based on monitored traffic behaviors. Meanwhile, a middle attention module is introduced between encoder and decoder to model the relationship between historical and future timesteps of traffic, thus it can alleviate the error accumulation and improve accuracy. The experiments based on both real-life dataset as well as synthetically generated traffic traces show that FlowDiviner can effectively obtain temporal and spatial correlation from the network historical traffic data, and the test results of all metrics are significantly improved from the baseline schemes.