Unsupervised Variable-Level Adversarial Representation Learning for Multivariate Time Series

Multivariable time series (MTS) representation learning is very helpful for downstream MTS analysis tasks. Existing representation models can accurately capture the individual features of each variable, but it is difficult to capture the correlation and difference between variables. To solve this issue, an unsupervised variable-level adversarial representation learning model is developed for MTS, called VATSRL. Firstly, our model regards multi-variate feature learning (overall features) and single-variate feature learning (single features) as two independent processes. Next, a novel “aggregation-decomposition” mechanism is designed and integrated into two learning processes to simulate the strong correlation between overall features and single features. Finally, to simulate the difference of two learning processes, the adversarial learning mechanism is integrated into the two learning processes to get more robust of representation vectors. Extensive experiments on three downstream tasks and thirteen public datasets demonstrate the excellent performance of our model.