DPGNet: Modeling Dynamic Graphs and Complex Temporal Patterns for Spatiotemporal Forecasting

Xiaobao Song, Hao Wang, Hongbo Qiu, Yuxin HE, Yanshan Li

Keywords: spatiotemporal forecasting, dynamic graph learning, complex temporal pattern, plug-and-play module

TL;DR: An efficient spatiotemporal forecasting model capable of capturing dynamic graphs and complex temporal patterns

Abstract:

Spatiotemporal prediction plays a crucial role in various domains, such as traffic management and weather forecasting.
Nevertheless, existing spatiotemporal prediction models remain limited in their ability to capture dynamic inter-node relationships and to comprehensively model complex temporal patterns.
To address these limitations, we propose the Dynamic Graph Prediction Network (DPGNet), which includes two key components: the Adaptive Graph Learner (AGL) and the Adaptive Season Learner (ASL).
AGL is a plug-and-play module that can effectively capture dynamic inter-node relationships while suppressing weak connections. %reducing the impact of noise.
ASL can model the complex temporal patterns of nodes and construct graph structures across different temporal patterns and time scales.
We conduct extensive experiments on five real-world spatiotemporal forecasting datasets, and the results demonstrate that DPGNet outperforms many state-of-the-art methods in both effectiveness and efficiency.
Our code is accessible at \url{https://anonymous.4open.science/r/DPGNet}.

Primary Area: learning on time series and dynamical systems

Submission Number: 2314