A unified multi-subgraph pre-training framework for spatio-temporal graph

Abstract

Spatio-temporal graph (STG) learning has shown great potential in capturing complex spatio-temporal dependencies and has achieved significant success in various fields such as traffic flow prediction, climate forecasting, and epidemiological spread research. By learning general features from spatio-temporal graphs, pre-trained graph models can capture hidden semantic information in the data, thereby enhancing the learning effect of downstream tasks and improving overall model performance. However, most existing spatio-temporal graph learning methods use the entire graph for training, which may not fully capture local structure and feature information. In addition, existing methods usually adopt sequence modeling techniques without fully considering the time decay effect, i.e., the need to apply decaying attention to distant time steps. To address these issues, this paper proposes a unified dual-phase multi-subgraph pre-training spatio-temporal graph framework (UMSST). Specifically, in the first phase, the framework learns the global representation of the spatio-temporal graph and locates key graph nodes, while learning the “unit representations” of these key nodes. In the second phase, multiple spatio-temporal subgraphs are constructed based on these “unit representations” to further capture the implicit encoding information of more general features around the corresponding subgraphs, thereby helping the model make full use of general features. Experimental results on real datasets show that the proposed pre-trained spatio-temporal graph framework significantly improves the performance of downstream tasks and demonstrates its effectiveness in comparison with recent strong baseline models.