Dynamic Spatial-Temporal Heterogeneous Graph Convolutional Network for Traffic Prediction

Accurate and effective traffic prediction has been a non-negligible and challenging task in the field of intelligent transportation, which plays a very important role in traffic management and route planning. Although numerous spatial-temporalodeling approaches exist to address this problem, they cannot accurately capture the flexible spatial - temporal dependence of traffic data and the unique spatial heterogeneity information of the traffic network. To overcome the current limitations, a new dynamic spatial-temporal heterogeneous graph convolution model is proposed in this paper. Specifically, to effectively model complex spatial correlation, we design a method combining dynamic adjacency matrix and a spatial heterogeneity matrix, to fully consider the potential spatial heterogeneity information of traffic data. In addition, this paper makes use of gated fusion with self-adaptive fusion temporal dependence and spatial dependence to make it more pertinent. Further, we have done a lot of experiments on two types of real traffic data sets. The results show the superiority and stability of the model for long-term prediction while outperforming other baseline methods.