DDCG: Dual-granularity Dual-domain Collaborative Graph Neural Networks for Time Series Forecasting

Multivariate time series forecasting has been widely applied in domains such as finance, traffic, energy, and healthcare. However, real-world time series often exhibit complex inter-channel dependencies. We propose DDCG, a Dual-granularity Dual-domain Collaborative Graph Neural Network, which models temporal structures at both global and patch levels. At the global scale, DDCG uses a multi-scale frequency-domain Mixture-of-Experts to enhance spectral representations and suppress noise, capturing the overall temporal characteristics and feature patterns of the series. At the patch scale, it employs a graph-based patch-level channel–temporal interaction mechanism to share channel graphs via a smooth-sharing strategy, effectively reducing computational complexity while preserving modeling precision. A fusion module integrates representations across both granularities, achieving cross-scale time–frequency collaboration and unified representation learning. Experiments on multiple real-world datasets show that DDCG achieves competitive performance, demonstrating its effectiveness and generality for complex multivariate time series modeling.