Fang, Shen (Zhejiang Lab), Liu, Hongyan (Zhejiang Lab), Yu, Chengcheng (Tongji University), Xie, Tian (Zhejiang Lab), Hua, Wei (Zhejiang Lab)

Revisiting Spatio-Temporal Forecasting: Feature Propagation Carry More Weights Than How They Do

Scheduled for presentation during the Regular Session "Traffic prediction and estimation I" (WeAT6), Wednesday, September 25, 2024, 10:30−10:50, Salon 14

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on December 26, 2024

Abstract

Spatio-temporal forecasting, a prominent research focus for understanding the dynamics of data flow in various domains, has recently extended its significance to traffic prediction as a notable application. Considering topology of data flow, existing methods mainly utilize Graph Convolutional Networks (GCNs), where graph construction is the basic concern and generally determines how data features are propagated. However, through extensive investigations and theoretical evidence, it is revealed that graph construction strategy, which is usually regarded as the key step to success, has actually provided very little benefit, while the presence of feature propagation itself on spatio-temporal domain is more significant, i.e., feature propagation carry more weights than how they do. Thus, by making slight refinements of a feature normalization method, we propose a Spatio-Temporal Propagation (STP) module, which does not require intervention of a specific graph structure, yet simple and effective. Various experiments on public datasets verify that the proposed STP module is an on-the-shelf tool that can be accessed to the end of current models or even replace GCNs as an alternative for capturing spatio-temporal features, while achieving better predictions. All the source codes are open accessed on GitHub.