STAN: A Spatio-Temporal Attention Network for Traffic Forecasting

Abstract

Traffic prediction plays a vital role in intelligent transportation systems, such as flow control and route planning. As a classical spatial-temporal prediction task, timely accurate traffic prediction is challenging due to complicated spatial-temporal correlation in traffic data. Traffic data contains short-term and long-term temporal sequential correlations, as well as local and global spatial correlations. In addition, traffic patterns at different time steps and the distribution of traffic data at various nodes may differ, so models need to have the ability to capture these temporal and spatial heterogeneities. To solve these issues, we propose a novel traffic prediction model, called Spatio-Temporal Attention Network (STAN). It models short-term and long-term temporal correlations using Gate Recurrent Unit (GRU) and Heterogeneity-Aware Temporal Self-Attention (HA-TSA) mechanism, respectively. It further models local and global spatial correlations using Graph Attention Network (GAT) and Heterogeneity-Aware Spatial Self-Attention (HA-SSA) mechanism, respectively. Based on traditional self-attention mechanism, HA-TSA and HA-SSA further learn different parameters for different nodes and different time steps to capture temporal and spatial heterogeneities. Besides, in order to extract the periodicity of traffic data, time stamps are concatenated to the original traffic features as auxiliary information. STAN also adds a centrality encoding based on node's degree to measure the importance of a node in the traffic network, thus helping to discover similar traffic patterns between nodes. Experimental results on two benchmark traffic datasets show that compared with the state-of-the-art baseline method, Mean Absolute Errors (MAEs) of STAN decrease by 7.48% and 7.77%, and Root Mean Squared Errors (RMSEs) of STAN decrease by 1.89% and 5.19%, respectively.