TCHA: Contrastive Learning via Hybrid Granularities and Adaptive Sampling for Multivariate Time Series

Learning discriminative representations from un-labeled data is essential for multivariate time series (MTS) analysis. Recently, contrastive learning algorithms have achieved remarkable success in MTS representation learning. However, these algorithms mostly contrast samples on a single subsequence or timestamp granularity without considering the contribution of different granularities, while the comparison among times-tamps, subsequences, or instances are all the gist of time series contrast and dominate the constrained mapping relationships. Additionally, existing unsupervised time series contrast methods often lack an accurate division method of positive and negative samples, resulting in many false negative samples during the contrast process and restricting the effect of contrast. Hence, in this work, we propose TCHA, an unsupervised contrastive learning framework unifying hybrid granularities and adaptive sampling for MTS. TCHA contrasts MTS at three granularities: timestamp, subsequence, and instance, effectively incorporating the semantic information of different scales in MTS. Moreover, we introduce an adaptive sampling method using similarity measurement to reduce false negative samples. Finally, the model is tested on 30 UEA datasets, and the experimental results demonstrate the superiority of TCHA over previous competitive baselines.