SSSLN:Multivariate Time Series Forecasting via Collaborative Dynamic Graph Learning.
Journal:
Neural networks : the official journal of the International Neural Network Society
Published Date:
Apr 26, 2025
Abstract
Multivariate time series (MTS) forecasting has achieved notable progress through graph modeling. However, existing approaches often face two key challenges. First, traditional dynamic graph learning (DGL) methods typically maintain dynamic graphs directly on the original MTS data, ignoring the distinct temporal patterns (e.g., trend and seasonality) that arise at different scales. Second, existing DGL methods often fail to capture interdependencies between graph structures evolving at multiple temporal scales, which can limit their ability to model complex spatio-temporal dynamics. To address these limitations, we propose the Synergistic Spatial Semantic Learning Network (SSSLN), a novel DGL-based framework for MTS forecasting. Specifically, our method (1) decomposes MTS data into trend and seasonal components and constructs separate dynamic graphs for each component to capture their unique temporal patterns, and (2) introduces a synergistic spatial semantic learning paradigm that explicitly aligns the spatial semantics of these graphs to guide their evolution. This design enables our model to fully exploit multi-scale information and uncover spatio-temporal dependencies across different scales. Extensive experiments on six real-world datasets across various domains demonstrate the effectiveness of our approach (e.g., the improvement in RSE score reaching as high as 5.913%).
