Reproduce “Revisiting Multivariate Time Series Forecasting with Missing Values” on OpenResearch

About this paper

Jie Yang, Yifan Hu, Kexin Zhang, Luyang Niu, Philip S. Yu, Kaize Ding

The CRIB framework, developed by researchers from UIC, Tsinghua, Northwestern, and Tongji Universities, proposes a direct prediction paradigm for multivariate time series forecasting with missing values. It demonstrated an average 13% reduction in MAE on PEMS-BAY and over 18% on ETTh1 compared to state-of-the-art baselines, by avoiding imputation and learning robust representations from partially observed data.

Key insights

• Unsupervised imputation without ground truth for missing values can actively harm forecasting accuracy by distorting data distributions and variate correlations.
• Direct prediction from partially observed data, guided by information-theoretic principles like Information Bottleneck, can yield superior and more robust forecasts than imputation-based methods.
• Combining patching for local features, unified-variate attention for global dependencies, and consistency regularization for stability is effective in learning robust representations from sparse multivariate time series.

Summary by alphaXiv.