LTCR: Long Temporal Characteristic Reconstruction for Segmentation in Contrastive Learning

Yang He, Yuhan Wu, Junru Zhang, Yabo Dong

Abstract: Contrastive learning is extensively employed for time series representation learning. Existing contrastive learning methods often split long real-world time series records into short samples along the temporal dimension to mitigate the explosive growth in storage and enhance the effectiveness of representations. However, the segmentation in contrastive learning may result in the loss of long characteristics, such as periodicities and trends of the original time series. To address this issue, we propose LTCR to reconstruct these long temporal characteristics after the segmentation. There are two key components of LTCR. On the one hand, we propose a novel definition of positive and negative pairs for contrastive learning. We treat the samples separated by integer multiples of a long periodicity from the same record as positive pairs and the other samples from the same record as negative samples to capture the periodic similarity, where the periodicity is calculated by the analysis in the frequency domain. On the other hand, we apply relative position prediction in the representation space to mine the trend characteristics of the original records. Comprehensive experiments have been carried out on three real-world time-series datasets, AD, TDBrain, and NerveDamage, which have been widely used in the field of time series representation. The results have demonstrated that LTCR significantly enhances the effectiveness of current contrastive learning methods, working as a plug-and-play plugin. We release the code and datasets at https://anonymous.4open.science/r/LTCR-76D5.