Graph Transformer Attention Networks for Traffic Flow Prediction

As a significant function of an intelligent transportation system (ITS), accurate traffic flow prediction plays an important role in traffic management. However, the nonlinear and complex characteristics of traffic flow makes the traditional approaches unable to achieve satisfactory prediction performance. Although the existing methods based on deep learning models have improved the accuracy of traffic flow prediction, they still ignore some potential features of traffic data. In this paper, we propose a novel network model called Graph Transformer Attention Network (GTAN) for traffic flow prediction. Firstly, we combine Transformer with Graph Convolutional Network (GCN) to discover the relationships between nodes by considering the information of all the positions of nodes in the traffic network and efficiently extract the spatial and temporal features of traffic flow. Secondly, to better exploit the flow characteristic of each node, we assign learnable parameters for each node separately to learn its unique variance. Finally, the experimental results on two real datasets show that the proposed model can achieve better prediction performance and less computation complexity compared with the other existing methods.