Dual-GT: Dual-scale Spatial Dependency for Grid-based Traffic Flow Prediction

As a crucial component of intelligent transportation system, grid-based traffic flow prediction has gained an extensive application in the smart city. In general, it segments the city into equal regions and aims to accurately predict the flow for each of them. Despite the progress achieved, existing works solely rely on the static spatial connections to capture the grids dependencies, neglecting the dynamic semantic dependencies over different time scales, which inevitably results in a degradation of performance. To this end, we propose a novel Dual-scale spatial dependency for Grid-based Traffic flow prediction model, called Dual-GT. Specifically, we design a dual-scale spatial dependency learning mechanism to capture the grid dependencies from both long-term and short-term time scales. In detail, we learn the long-term dependencies through the Dynamic Time Warping algorithm. Furthermore, we innovatively characterize the short-term semantic dependencies by clustering the flow data series, and design an adaptive transfer method to efficiently integrate the short-term dependencies with the learned long-term ones. Extensive experiments on five real-world public traffic datasets verify the superiority of our approach. Additionally, we visualize the spatial dependency learned from long-term and short-term traffic flows to further show the effectiveness and interpretability of our Dual-GT model.