Noise-tolerant universal representation learning for multivariate time series from global-to-local perspective

Representation learning for multivariate time series (MTS) has shown great potential in various analysis tasks. However, most existing representation models are designed for a certain task, such as forecasting and classification, which may cause poor generality and incomplete feature extraction. Moreover, these models are often sensitive to noise, which also affects the performance. To address these issues, an unsupervised, noise-tolerant universal representation learning model, namely TSG2L, is proposed for multivariate time series from a global-to-local perspective. Inspired by the idea of “drawing the outline before filling in the details”, TSG2L adopts a global-to-local learning way instead of the traditional local-to-global way. Technically, TSG2L divides the representation learning process into two sequential stages: global feature learning (drawing outline) and local feature learning (filling detail). In the first stage, a noise-tolerant multi-scale global reconstruction network is designed to perform variable-independent global feature learning. In the second stage, a noise-tolerant “1+M” prediction network is developed to integrate global features and perform variable-related local feature learning. To the best of our knowledge, this is the first work to explore MTS representation learning from a global-to-local perspective. Extensive experiments on three analysis tasks and eighteen real-world datasets demonstrate that TSG2L outperforms several state-of-the-art models. The source code of TSG2L is available at https://github.com/infogroup502/TSG2L.