SDMixer: Sparse Dual-Mixer for Time Series Forecasting

Authors: Xiang Ao Date: 2026-02-27 Paper ID: openalex:2602.23581

Summary

The SDMixer model addresses challenges in multivariate time series forecasting, such as multi-scale characteristics and weak correlations, by employing a dual-stream framework. One stream operates in the frequency domain to capture global trends, while the other operates in the time domain to capture local dynamic features. A core component is a sparsity mechanism designed to aggressively filter noise and invalid information, thereby strengthening the accurate modeling of cross-variable dependencies. Experimental validation across several real-world datasets demonstrates that SDMixer sets a new state-of-the-art performance level.

Key Contributions

• Proposed the Sparse Dual-Mixer (SDMixer) framework, featuring dual streams for frequency and time domain feature extraction.
• Introduced a sparsity mechanism within the Mixer architecture to filter invalid information and enhance cross-variable dependency modeling.
• Achieved leading performance compared to existing models on multiple real-world time series forecasting benchmarks.

Limitations

The paper does not explicitly detail limitations, but a potential area for future work is investigating the optimal sparsity level determination.

Open Questions & Future Work

• dynamic-sparsity-in-time-frequency-decoupling

Datasets

• ETTh1
• Traffic

Limitations

The paper does not explicitly detail limitations, but a potential area for future work is investigating the optimal sparsity level determination.

Links

• Abstract
• PDF