A local global attention based spatiotemporal network for traffic flow forecasting.

Accurate traffic forecasting is critical to improving the safety, stability, and efficiency of intelligent transportation systems. Although many spatiotemporal analysis methods have been proposed, accurate traffic prediction still faces many challenges, for example, it is difficult for long-term predictions to model the dynamics of traffic data in temporal and spatial to capture the periodicity and spatial heterogeneity of traffic data. Most existing studies relieve this problem by discovering hidden spatiotemporal dependencies with graph neural networks and attention mechanisms. However, the period-related information between spatiotemporal sequences is not sufficiently considered in these models. Therefore, we propose a local global spatiotemporal attention network (LGA) to solve the above challenge. Specifically, we present a local spatial attention module to extract the spatial correlation of hourly, daily, and weekly periodic information. We propose a weight attention mechanism to assign different weights of the periodic feature extracted on local spatial attention. The local periodic temporal features are extracted through the local temporal attention we proposed. And we develop the global spatiotemporal attention module to extract the global spatiotemporal information of the entire time slice, which is more conducive to learning the periodic features of traffic data. The extensive experiments on four real-world datasets demonstrate the effectiveness to our proposed model.The code is publicly available on github 1 (github: https://github.com/lyc2580/LGAmodel.). [ABSTRACT FROM AUTHOR]