Mining the Graph Representation of Traffic Speed Data for Graph Convolutional Neural Network

Abstract

Deep learning algorithms are considered as the best-fit methods to deal with spatial-temporal attributes of short-term traffic predictions in recent years. Further, the graph-based Graph Convolutional Network (GCN) models are widely used to handle the spatial dependence of roads in urban networks. This paper aims to explore the spatial graph representation of urban networks for GCN models. Specifically, a data-driven spatial graph representation scheme is established to measure the complex non-linear relationships among roads, together with the local and non-local impacts of urban traffics. This spatial graph representation is then combined with Sequence to Sequence structure to present a multi-input and multi-output network-wide traffic prediction model (SGDE-S2S model). A sensitive test is carried out to select the optimal threshold value of the most relevant roads to every target road. Then the SGDE-S2S model and some other baseline models are tested on real-world traffic speed data of Chengdu, China. The experiment results confirm that the SGDE-S2S model can well capture the intrinsic relationships of roads without the need for topological adjacent information and performs the best in all multi-step predictions.