Abstract

Abstract Traffic forecasting is a challenging problem due to the complexity of jointly modeling spatio‐temporal dependencies at different scales. Recently, several hybrid deep learning models have been developed to capture such dependencies. These approaches typically utilize convolutional neural networks or graph neural networks (GNNs) to model spatial dependency and leverage recurrent neural networks (RNNs) to learn temporal dependency. However, RNNs are only able to capture sequential information in the time series, while being incapable of modeling their periodicity (e.g., weekly patterns). Moreover, RNNs are difficult to parallelize, making training and prediction less efficient. In this work we propose a novel deep learning architecture called Traffic Transformer to capture the continuity and periodicity of time series and to model spatial dependency. Our work takes inspiration from Google’s Transformer framework for machine translation. We conduct extensive experiments on two real‐world traffic data sets, and the results demonstrate that our model outperforms baseline models by a substantial margin.