Sequential Graph Neural Network for Urban Road Traffic Speed Prediction

Abstract

Accurate speed predictions for urban roads are highly important for traffic monitoring and route planning, and also help relieve the pressure of traffic congestion. Many existing studies on traffic speed prediction are based on convolutional neural networks, and these have primarily focused on capturing the spatial proximity among different road segments. However, the real cause of the spread of traffic congestion is the connectivity of these road segments, rather than their spatial proximity. This makes it very challenging to improve prediction accuracy. Using graph neural networks (GNNs), the connectivity of these road segments can be modeled as a graph in which the properties of road segments and the connections between them are embedded as the properties of the nodes and edges, respectively. This paper describes a novel approach that combines the advantages of sequence-to-sequence (Seq2Seq) models and GNNs. Specifically, the evolution of traffic conditions on road networks is modeled as a sequence of graphs. Thus, the proposed SeqGNN model represents both the inputs and outputs as graph sequences. Finally, the extensive experiments using real-world datasets demonstrate the effectiveness of our approach and its advantages over the state-of-the-art methods.