DyTimeNet: Dynamic cross-variable dependency network with sparse strategy for multivariate time series forecasting

Abstract

Most current research on multivariate time series focuses on extracting temporal patterns, often treating the data as univariate and neglecting interactions between channels. This work introduces variable-wise auxiliary information to facilitate forecasting. In reality, channel relationships are dynamic and sparse, with changing correlations over time, and fusing all channels may distort the underlying data structure. To address this, we propose DyTimeNet, a novel dual-path learning framework designed to capture dynamic sparse correlations between variables (CrossVBlock) and complex nonlinear relationships within time series (IntraVBlock). Specifically, we introduce an innovative DropFusion sparse strategy within CrossVBlock, which simulates multiple variable-wise correlations by dynamically selecting connections (edges) and dropping neurons. By doing so, it reduces unnecessary interference from irrelevant variables, forcing the network to avoid over-relying on specific edges or neurons, thereby adapting to varying patterns of channel correlations. Experiments on thirteen real-world datasets demonstrate that DyTimeNet significantly outperforms sophisticated Transformer-based models and even surpasses LLM-based models, especially in capturing better fine-grained pattern fitting in multivariate time series. Additionally, DyTimeNet uses only 1/4 of the parameter size and trains 7$\times$ faster than state-of-the-art models like iTransformer, while improving MSE by 10.13%. These improvements make DyTimeNet ideal for applications with limited computational resources or high-frequency demands, such as edge-side sensor networks.