DCANet: Dual-Path Complementary Attention Network for Multivariate Time Series Forecasting

Multivariate time series forecasting requires capturing both individual series patterns and interactions between series. Existing methods often struggle to balance these two aspects. Moreover, the widely used attention mechanism may over focus on certain parts of the series, increasing the risk of overfitting. To address these issues, we propose DCANet, a Dual-Path Complementary Attention Network that introduces a novel complementary attention mechanism to balance series-specific and cross-series modeling. DCANet has three key components: (1) Heads-Mixing, which reformulates each series’ embedding into multiple low-rank heads and applies self-attention across these heads to capture fine-grained within-series patterns; (2) Attention Decomposition, which divides attention weights into focused and complementary components, isolating the primary dependencies and the secondary (neglected) context; and (3) Complementary Attention Fusion, which exchanges the complementary information between a series-focused path and a channel-focused path, integrating context that standard attention would overlook. By fusing focused and complementary attention outputs, DCANet effectively captures both local series details and broader cross-series context. Experiments on real-world multivariate time series benchmarks show improved accuracy over recent Transformer-based and Decomposition-based models, demonstrate the benefits of our dual-path complementary attention approach.