MST-GNN: Graph Neural Network with Multi-Granularity in Space and Time for Traffic Prediction

Xinru Zhao, Wenhao Yu, Lin Jianfeng, Yifan Zhang, Jinlu Zhang, Huang Binqiang

Keywords: traffic flow prediction, graph neural network, multi-scale spatio-temporal structure

Abstract: Traffic flow prediction based on vehicle trajectories is a crucial aspect of Intelligent Transportation Systems (ITS). Deep learning approaches have recently been widely adopted to capture spatio-temporal correlations in traffic conditions, and have achieved superior performance compared to traditional methods. However, most existing studies focus on traffic prediction at a single spatial scale, usually corresponding to the road-segment level. According to the Hierarchy Theory, processes at different scales form a hierarchy of organization, and meaningful patterns may emerge at multiple levels of details. Presetting traffic data at an inappropriate scale can cause misunderstanding in features learning. In this paper, we propose a graph learning model called MST-GNN, which captures the comprehensive behaviors and dynamics of traffic patterns by fusing multi-scale features from both space and time perspectives. In ITS applications, users usually consider traffic conditions at the larger-scale regional level, and a prediction model must attend to multi-scale application requirements. Moreover, the structure of multiple granularities in time series can fully unleash the potential of different temporal scales in learning dynamic traffic pattern features. We inject the multi-scale spatio-temporal structure into a graph neural architecture with a tailored fusion module. Our model achieves state-of-the-art accuracy prediction results on two traffic benchmark datasets.

Primary Area: general machine learning (i.e., none of the above)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2360