A dynamical spatial-temporal graph neural network for traffic demand prediction

Traffic demand prediction is significant and practical in the resource scheduling of transportation application systems. Meanwhile, it remains a challenging topic due to the complexities of contextual effects and the highly dynamic nature of demand. Many works based on graph neural network (GNN) have recently been proposed to cope with this task. However, most previous studies treat the spatial dependence as a static graph, and their inference mechanism lacks interpretability. To address the issues, a Dynamical Spatial-Temporal Graph Neural Network model (DSTGNN) is proposed in this paper. DSTGNN has two critical phases: (1) Creating a spatial dependence graph. To capture the dynamical relationship, we propose building a spatial graph based on the stability of node's spatial dependence. (2) Inferring intensity. We model the changing demand process using the inhomogeneous Poisson process, which addresses the interpretability issue, and build a spatial-temporal embedding network to infer the intensity. Specifically, the spatial-temporal embedding network integrates the diffusion convolution neural network (DCNN) and a modified transformer. Extensive experiments are carried out on two real data sets, and the experimental results demonstrate that the performance of DSTGNN outperforms the state-of-the-art models on traffic demand prediction. (C) 2022 Elsevier Inc. All rights reserved.

Automated summary

Traffic demand prediction is significant and practical in resource scheduling. A Dynamical Spatial-Temporal Graph Neural Network model (DSTGNN) is proposed in this paper to address the challenges of contextual effects and dynamic demand. DSTGNN creates a spatial dependence graph and infers intensity using an inhomogeneous Poisson process. Extensive experiments demonstrate its superior performance in traffic demand prediction.