Multivariate Correlation Self-Distillation Transformer for Time Series Forecasting With Incomplete Data

Abstract

Multivariate time series forecasting estimates future development by capturing variable relationships and constructing temporal regular, which is widely used in many scenarios, including industrial production, economic development, and disease prediction. Although the existing deep learning methods have achieved impressive results in multivariate time series forecasting tasks, the existing methods only emphasize the prediction performance and ignore the widespread issue of missing data in the real world. This article proposes a robust multivariate correlation self-distillation Transformer framework for incomplete time series data forecasting. The proposed method first decouples the interinference of historical series and the exter-inference of future series into two stages. The first stage focuses on the reconstruction of historical series, while the second stage focuses on the prediction of future series. Then, a novel multivariate correlation Transformer is designed as the basic component of the network, which can perform feature inference from both multivariate relationships and single-variate temporal regular. Finally, a variable correlation self-distillation method is proposed to self-distill the more complete variable relationship from the exter-inference stage to the interinference stage. The proposed method is verified on eight real-world datasets, and both qualitative and quantitative results show that the proposed method has good performance.