DHINet: decomposed hybrid interaction network

Multivariate time series forecasting is a critical task applied across numerous domains, including meteorology, finance, and traffic management. However, the inherent non-stationary and stochastic nature of real-world time series data presents significant challenges for accurate modeling and prediction. Many recent forecasting methods adopt channel independence strategies to mitigate issues arising from data distribution shifts. While beneficial in some respects, this approach can lead to a loss of mutual information between variables, thereby limiting potential performance gains. Furthermore, prevailing studies often employ trend decomposition techniques that focus primarily on intra-component interactions (i.e., within trend and seasonal terms) while overlooking the crucial interdependencies between the trend and seasonal components themselves. To address these limitations, we introduce the Decomposed Hybrid Interaction Network (DHINet), a novel hybrid dual-path architecture integrating Convolutional Neural Networks (CNNs) and Fourier Attention. DHINet features two key components: a Decomposed Hybrid Interaction (DHI) module and a difference-based compensation prediction module(DCP). The network is specifically engineered to explicitly model seasonal-trend relationships by strategically leveraging both channel mixing and channel independence. This facilitates a structured channel information flow from seasonal to trend components and vice versa. Additionally, DHINet utilizes differencing alongside standard projection methods for future window prediction, effectively reducing prediction lag. As a result, DHINet achieves state-of-the-art performance in both long-term and short-term forecasting scenarios.