Spatio-temporal graph attention networks for traffic prediction

Abstract

The constraints of road network topology and dynamically changing traffic states over time make the task of traffic flow prediction extremely challenging. Most existing methods use CNNs or GCNs to capture spatial correlation. However, convolution operator-based methods are far from optimal in their ability to fuse node features and topology to adequately model spatial correlation. In order to model the spatio-temporal features of traffic flow more effectively, this paper proposes a traffic flow prediction model, the Spatio-Temporal Graph Attention Network (STGAN), which is based on graph attention mechanisms and residually connected gated recurrent units. Specifically, a graph attention mechanism and a random wandering mechanism are used to extract spatial features of the traffic network, and gated recurrent units with residual connections are used to extract temporal features. Experimental results on real-world public transportation datasets show that our approach not only yields state-of-the-art performance, but also exhibits competitive computational efficiency and improves the accuracy of traffic flow prediction.