UniMixer: Unified Patch-Wise and Global Inter-Series Dependency Modeling for Multivariate Time Series Forecasting

Multivariate time series forecasting is crucial in domains such as finance, energy, and transportation, requiring effective modeling of temporal dynamics and inter-variable dependencies. Existing methods often emphasize either local or global dependency modeling but struggle to seamlessly integrate both and accurately capture complex inter-variable relationships. To address these challenges, we propose UniMixer, a unified framework designed to effectively model dependencies within multivariate time series data. UniMixer integrates three key components: the Patch-Wise Mixer, which captures local temporal patterns; the Global Context Enhancer, which models long-range inter-series relationships; and the Correlation Token Mapper, which explicitly encodes inter-variable correlations. This design achieves a balance between local detail preservation and global dependency understanding, demonstrating strong performance across various forecasting tasks. Extensive experiments conducted on eight real-world multivariate datasets demonstrate that UniMixer achieves superior performance compared to state-of-the-art (SOTA) methods on most datasets, highlighting its effectiveness and adaptability. By unifying local and global dependency modeling, UniMixer establishes a robust foundation for advancing multivariate time series forecasting, while offering insights into inter-variable relationships. Our code is available at: https://github.com/CYD-y/UniMixer .