TYRE: A dynamic graph model for traffic prediction

In this paper, we study the problem of traffic forecasting, which aims to predict the future traffic state of the road network. One key challenge is that the previous approaches lack discussion of capturing temporal dependencies, as well as spatial dependencies among locations in the traffic network. In addition, the long-term traffic prediction is not satisfied. In this paper, we propose a raffic dnamic gaph modl — — which is composed of a Graph Convolutional Network with Gated and Attention mechanisms. TYRE can learn the ‘importance’ of all adjacent and distant locations, control the aggregation of adjacent and distant neighbourhood information, and learn the temporal dependencies to support long effective historical sizes. We demonstrate the validity and effectiveness of our approach on two different traffic datasets (i.e., PeMSD4 and PeMSD8). The result shows that compared to the related approaches, our model that captures temporal and spatial dependence yields substantially improved performance. When predicting traffic conditions for the next 120 min, on PeMSD8 dataset, our model shows almost 6.6% RMSE improvement, 10.9% MAE improvement, and 2.1% MAPE improvement over the previous state of the art. All source codes of this work will be publicly available at