Spatial-Temporal Multiscale Fusion Graph Neural Network for Traffic Flow Prediction

Abstract

Traffic flow prediction is an important task for building an intelligent traffic system. However, it is a challenging task because of the spatial-temporal heterogeneity of traffic data. Existing methods usually capture spatial and temporal dependencies separately through complex mechanisms, ignoring the spatial-temporal heterogeneity of traffic data. Besides, they only capture the connectivity of adjacent time series and the spatial connectivity of traffic nodes, lacking the capture of complex spatial-temporal dependencies. To overcome these problems, we propose a spatial-temporal multiscale fusion graph neural network (MFGNN) model for traffic flow prediction. MFGNN generates time series similarity graph and regional clustering graph in a data-driven manner, capturing the spatial-temporal dependencies of traffic nodes from multiple scales. MFGNN connects consecutive time steps to generate the fusion graph, which contains the spatial connectivity, regional clustering relationship, time series similarity, and adjacent time series connectivity of traffic nodes. By performing the two-layer lightweight graph convolution operations on the fusion graph, MFGNN can capture spatial and temporal dependencies synchronously to extract the spatial-temporal heterogeneity of traffic data. Extensive experiments on three real-world datasets show that our model achieves state-of-the-art performance consistently than other baselines.