Spatio-Temporal Graph Attention Network for Traffic Prediction

Abstract

Timely and accurate traffic prediction plays an important role in urban traffic management and planning. It forecasts the most likely future traffic volume based on the patterns and characteristics of historical time series data. Generally, a complete road network is composed of multiple urban roads, which can influence each other dynamically. Most of the existing traffic prediction methods rely on the specific network structure to compute the fixed Laplacian matrix for mining spatial dependencies. In reality, the correlations between roads are not fixed, but change over time due to the travel trends and routing habits of citizens. Moreover, unexpected events such as traffic accidents and traffic jams can happen, which also results in the change of spatial dependencies. Under these circumstances, traditional methods are incapable of accurate traffic forecasting with time-evolving spatial correlations. In this paper, we proposed a new model named Spatio-Temporal Graph Attention Network (STGAT) for highway traffic prediction. Gated CNNs with standard 1D convolution is applied for temporal modelling with high efficiency. Graph Attention Network (GAT) is utilized for mining the dynamic spatial dependencies hidden in the traffic data. Experimental results on two real-world datasets validate the effectiveness of the proposed model.