Dynamic graph structure learning for multivariate time series forecasting

• Published in: Pattern recognition Vol. 138; p. 109423
• Main Authors: Li, Zhuo Lin, Zhang, Gao Wei, Yu, Jie, Xu, Ling Yu
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.06.2023
• Subjects: Dynamic spatio-temporal dependencies; Graph neural networks; Graph structure learning; Long- and short-term patterns; Multivariate time series forecasting; Dynamic spatio-temporal dependencies; Graph structure learning; Graph neural networks; Multivariate time series forecasting; Long- and short-term patterns
• ISSN: 0031-3203; 1873-5142
• DOI: 10.1016/j.patcog.2023.109423

•We formalize the problem of dependencies in multivariate time series data being a mixture of long- and short-term patterns.•We propose a novel graph learning-neural network to model long- and short-term patterns in data without any priori knowledge.•The propose dynamic graph learning method can capture dynamic spatio-temporal dependencies in short-term patterns.•We show the effectiveness on six public datasets and analyze the learned graph structures. Multivariate time series forecasting is a challenging task because the dynamic spatio-temporal dependencies between variables are a combination of multiple unknown association patterns. Existing graph neural networks typically model multivariate relationships with a predefined spatial graph or a learned fixed adjacency graph, which fails to handle the aforementioned challenges. In this study, we decompose association patterns into stable long-term and dynamic short-term patterns and propose a novel framework, named the static and dynamic graph learning network (SDGL), for modeling unknown patterns. Our approach infers two types of graph structures, from the data simultaneously: static and dynamic graphs. A static graph is developed to capture the fixed long-term pattern via node embedding, and we leverage graph regularity to control its learning direction. Dynamic graphs, which are time-varying matrices based on changing node-level features, are used to model dynamic dependencies over the short term. To effectively capture local dynamic patterns, we integrate the learned long-term pattern as an inductive bias. Experiments on six benchmark datasets show the state-of-the-art performance of our method. Analysis of the learned graphs reveals that the model succeeds in modeling dynamic spatio-temporal dependencies.