IGAGCN: Information geometry and attention-based spatiotemporal graph convolutional networks for traffic flow prediction

Abstract

In this study, a novel spatiotemporal graph convolutional networks model is proposed for traffic flow prediction in urban road networks by fully considering an information geometry approach and attention-based mechanism. Accurate traffic flow prediction in real urban road networks is challenging due to the presence of dynamic spatiotemporal data and external factors in the urban environment. Moreover, the dynamic spatial and temporal dependencies of urban traffic flow data are very important for predicting traffic flow, and it has been shown that a recent attention mechanism has a relatively good ability to capture these dynamic dependencies, which are not fully considered by most existing algorithms. Therefore, in the novel model abbreviated as IGAGCN, the information geometry method is utilized to determine the dynamic data distribution difference between different sensors. The attention mechanism is employed with the information geometry method, in which a matrix is derived by analyzing the distributions of sensor data, and the spatiotemporal dynamic connections in traffic flow data features are better at capturing the spatial dependencies of traffic between different sensors in urban road networks. Furthermore, a parallel sub-model architecture is proposed to consider long time spans, where each dilated causal convolution sub-model is applied to short time spans. Two well-known data sets were employed to demonstrate that our proposed method obtains better performance and is better at capturing the dynamic spatial dependencies of traffic than the existing only-attention-based models. In addition a real-world urban road network in Shenzhen, China, was studied to test and verify the proposed model.