DSTN: Early Spatio-Temporal Forecasting with Dynamic Propagation

Shuhan Wu, Wei Shao, Yufan Kang

Keywords: Spatio-Temporal Data, Graph Neural Network, Early Detection

Abstract: In most spatio-temporal prediction tasks, the timeliness of predictions is more critical than their accuracy. For instance, in tasks such as crime prediction, traffic congestion forecasting, and wildfire early warning, waiting longer to gather additional information may improve prediction accuracy, but it does not provide enough preparation time for subsequent actions, rendering the precise predictions valueless. Therefore, balancing between prediction timeliness and accuracy is essential for such tasks. In this paper, we propose an adaptive early spatio-temporal prediction model with a dynamic propagation matrix (DSTN), which captures causal relationships between nodes to enhance prediction timeliness while maintaining accuracy. Our model makes the following contributions: (1) Exploiting the similar long-term patterns of node signals for early prediction. (2) Proposing the concept of Asynchronous Spatio-temporal Causal Frame Pair to effectively capture the spatio-temporal causal relationships between different nodes. (3) Constructing a dynamic propagation matrix to filter out irrelevant information for early prediction. Experimental results on four large-scale real-world datasets demonstrate that the performance of our proposed DSTN model generally outperforms all baselines. The source code is available at https://anonymous.4open.science/r/DSTN-DB49.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 4115