Multivariate Time Series Forecasting with Dynamic Graph Neural ODEs

Multivariate time series forecasting has long received significant attention in real-world applications, such as energy consumption and traffic prediction. While recent methods demonstrate good abilities, they have three fundamental limitations. (i) Discrete neural architectures: Interlacing individually parameterized spatial temporal blocks to encode rich underlying patterns leads discontinuous latent state trajectories higher numerical errors. (ii) High complexity: approaches complicate models with dedicated designs redundant parameters, leading computational memory overheads. (iii) Reliance on graph priors: Relying predefined static structures limits their effectiveness practicability applications. In this paper, we address all the above limitations by proposing a continuous model forecast $\textbf{M}$ultivariate $\textbf{T}$ime dynamic $\textbf{G}$raph $\textbf{O}$rdinary $\textbf{D}$ifferential $\textbf{E}$quations ($\texttt{MTGODE}$). Specifically, first abstract multivariate into graphs time-evolving node features unknown structures. Then, design solve ODE complement missing topologies unify both message passing, allowing deeper propagation fine-grained information aggregation characterize stable precise spatial-temporal dynamics. Our experiments superiorities of $\texttt{MTGODE}$ from various perspectives five benchmark datasets.