Modeling of spatial–temporal dynamic dependency in traffic data to predict its evolution

Abstract

Accurate traffic forecasting enables efficient traffic management. However, traffic prediction is a challenging task as the transportation system itself presents complex dynamic characteristics due to the complex interactions of multiple agents (such as randomly mixed vehicles with different mechanical characteristics, drivers with diverse driving habits, dynamic traffic lights, and occasional traffic events). To model the dynamic spatial–temporal dependencies caused by these complex interactions, we propose a new Multi-Head self-Attention based Spatial–Temporal Information Graph Convolutional Network (MH-ASTIGCN). Firstly, we propose a data-driven strategy for generating temporal information graphs to capture the spatial correlation that cannot be fully obtained by static spatial adjacency graphs. Secondly, we design a novel spatial–temporal attention generation module (STAGM) to capture the complex dependencies among traffic nodes via multi-subspace learning, where the features in graph convolution are adaptively aggregated by dynamically adjusting each term of the Chebyshev polynomial. Thirdly, we propose a spatial–temporal feature aggregation module (STFAM), which can effectively extract multi-order neighborhood information in space and fuse multi-receptive fields from multi-subspace features in time. Finally, we integrate our improved STAGM and STFAM into the graph convolutional neural network to predict the traffic flow.Experiments on three real-world data sets show that our proposed method outperforms baseline methods, especially in long-term prediction. More importantly, the dynamic dependency characteristics among traffic nodes have been visualized, which makes a substantial contribution to Artificial Intelligence in modeling and understanding the complex dynamic dependencies among traffic nodes. The codes are available at https://github.com/SYLan2019/MH-ASTIGCN .