Spatiotemporal Representation Learning on Time Series with Dynamic Graph ODEs
Spatiotemporal representation learning on multivariate time series has received tremendous attention in forecasting traffic and energy data. Recent works either rely on complicated discrete neural architectures or graph priors, hindering their effectiveness and applications in the real world. In this paper, inspired by neural ordinary differential equations and graph structure learning, we propose a fully continuous model named Dynamic Graph ODE (DyG-ODE) to capture both long-range spatial and temporal dependencies to learn expressive representations on arbitrary multivariate time series data without being restricted by rigid preconditions (e.g., graph priors). For modeling the continuous dynamics of spatiotemporal clues, we design a simple yet powerful dynamic graph ODE by coupling the proposed spatial and temporal ODEs, which not only allows the model to obtain infinite spatial and temporal receptive fields but also reduces numerical errors and model complexity significantly. Our empirical evaluations demonstrate the superior effectiveness and efficiency of DyG-ODE on a number of benchmark datasets.
PDF Abstract
