DynGCN: Capturing Dynamic Correlation with Message Passing

Chen Shao, Yue Wang, Zhanbo Huang, Benjamin Schäfer, Tobias Käfer, Danai Koutra

Keywords: Multivariante Time Series Forecasting, Correlation Modeling, Graph Neural Network

TL;DR: SeriesGCN is a novel GNN framework that leverages high-order covariance modeling and dual graph aggregation to improve multivariate time series forecasting under both static and dynamic correlation scenarios, outperforming state-of-the-art baselines.

Abstract: A recent approach to modeling multivariate time series is to represent them as a graph, with time series as nodes and pairwise temporal correlations as edges. Advances in Graph Neural Networks (GNNs) have shown strong performance in multivariate time series forecasting by assuming a static graph topology and ag- gregating information from neighboring time series based on their correlations. In this work, we investigate the representational power of GNNs for short- and long-term forecasting under both static and dynamic correlation scenarios, i.e., when pairwise correlations remain fixed or evolve over time. We show that many popular GNNs generalize poorly in these settings and are even outperformed by structure-agnostic baselines. To address these limitations, we propose DYNGCN, a novel GNN framework enhanced by two theoretically justified designs: (D1) high-order moment based message passing and (D2) static and dynamic propa- gation seperation. These components improve learning under dynamic correla- tions while preserving robustness under static scenarios. Extensive experiments on synthetic and real-world benchmarks demonstrate that DYNGCN achieves up to 23.25% and 23.08% performance gains over state-of-the-art baselines.

Primary Area: learning on time series and dynamical systems

Submission Number: 24180