Reproduce “DualWeaver: Synergistic Feature Weaving Surrogates for Multivariate Forecasting with Univariate Time Series Foundation Models” on OpenResearch

About this paper

Jinpeng Li, Zhongyi Pei, Huaze Xue, Bojian Zheng, Chen Wang, Jianmin Wang

DualWeaver, developed by researchers at Tsinghua University and Tencent, introduces a framework to adapt univariate time series foundation models (Uni-TSFMs) for multivariate forecasting. It constructs dual learnable surrogate series to capture cross-variable dependencies and enables non-parametric prediction reconstruction, leading to consistent performance improvements, such as an 8.95% average MSE reduction for TimerXL, and enhanced robustness against existing adaptation methods.

Key insights

• Transforming the multivariate forecasting problem into predicting a pair of 'dual surrogates' allows Uni-TSFMs to implicitly learn complex cross-variable dependencies without architectural modifications.
• Deriving final predictions through non-parametric reconstruction from the dual surrogates ensures that the adapted model's capacity is fully dedicated to learning prediction-friendly dependencies, rather than auxiliary reconstruction.
• Integrating a theoretically derived error bound as a regularization term is crucial for maintaining training stability and ensuring the adapted model's error bound remains competitive with baseline methods.

Summary by alphaXiv.