Spatio-Temporal Consistency for Multivariate Time-Series Representation Learning

Abstract

Label sparsity in multivariate time series (MTS) makes using label information for practical applications challenging. Thus, unsupervised representation learning methods have gained attention to learn effective representations suitable for various MTS tasks without relying on labels. Recently, contrastive learning has emerged as a promising approach to generate robust representations by capturing underlying MTS information. However, the existing methods have some limitations, such as insufficient consideration of cross-variable relationships of MTS and high sensitivity to positive pairs. Therefore, we proposed a novel spatio-temporal contrastive representation learning method (STCR) designed to address these limitations. STCR focuses on learning robust representations by encouraging spatio-temporal consistency, which comprehensively considers spatial information as well as temporal dependencies in MTS. The results of extensive experiments on MTS classification and forecasting tasks demonstrate the efficacy of STCR in generating high-quality representations, achieving state-of-the-art performance on both tasks.