Learning Dynamics of Multi-level Spatiotemporal Graph Data for Traffic Flow Prediction

—Accurate and timely traffic flow prediction can effectively improve road network efficiency and alleviate a series of negative impacts caused by traffic congestion. Recent research has focused on exploring the spatiotemporal correlations of traffic data, and achieved some progress. However, due to the highly nonlinear and stochastic nature of traffic data, the most prediction methods can only learn the fixed physically connected relationships by predefined empirical matrices, but cannot fully capture the spatial correlation, especially the dynamic characteristics. In this paper, we propose a Multiscale Gated Spatiotemporal Graph Convolution Network (MS-GSTGCN). A self-learning adjacency matrix is constructed by the graph fusion module aggregating the information of the adaptive adjacency matrix and the innovative similarity adjacency matrix, which helps to capture the hidden features of the nodes and the dynamic spatial correlation. Meanwhile, a multilevel graph neural structure is designed to extract data features utilizing hierarchical framework, and a gating mechanism is introduced to minimize the accumulation of errors when the information propagates layer by layer, so as to enhance the model's ability to capture complex spatiotemporal correlations in long-term predictions. The experiments on two real world datasets show that it outperforms state-of-the-art related methods.