Multi-View Spatial-Temporal Adaptive Graph Convolutional Networks for Traffic Forecasting

Abstract

Traffic flow prediction is vital to traffic management and public safety, and it plays a very important role in a smart transportation system. Traffic flow data contains complex temporal relation, spatial relation, and spatial-temporal relation. In this paper, we propose a novel Multi-View Spatial-Temporal Adaptive Graph Convolutional Network (MVST-AGCN) for traffic flow prediction. MVST-AGCN is mainly composed of three independent modules, which construct temporal view, spatial view, and spatial-temporal view respectively. Specifically, MVSTAGCN integrates the structures of Residual Gated Linear Units (RGLU), Multi-layer Adaptive Graph Convolutional Network (MAGCN), and transformer, which correspond to temporal, spatial, and spatial-temporal views respectively. Different from the existing research, we use the multi-layer adaptive graph convolution based on the adaptive adjacency matrix to overcome the disadvantages of the traditional pre-defined adjacency matrix. Experiments on four real-world datasets demonstrate that the proposed MVST-AGCN model outperforms the state-of-the-art baselines.