ST-MoE: Spatio-Temporal Mixture of Experts for Multivariate Time Series Forecasting

Hanwen Liu, Yibing Zhang, Ximeng Wang, Bin Wang, Yanwei Yu

Abstract: Multivariate time series (MTS) forecasting aims to predict future values of multiple target variables, offering crucial decision-making insights in industries such as finance, transportation, and meteorology. The current mainstream research approach is to employ complicated deep-learning-based models to extract complex spatiotemporal features. However, recent research has shown that some structurally simple deep models can outperform complex structures such as GNN. Driven by this insight, this paper introduces a novel network architecture, called Spatial-Temporal Mixture-of-Experts (ST-MoE), which comprises a few simple experts and a MoE-based decision module for expert selection. The expert networks first merge spatial and temporal embeddings of original sequence data into a unified feature map, and the decision module then adaptively selects the part of specific experts for the final forecasting. We conduct comprehensive experiments on five public datasets, benchmarking ST-MoE against multiple baselines across the perspectives of prediction accuracy, model efficiency, and ensemble effectiveness, and demonstrate the superiority of the proposed ST-MoE.