# Learning Timestamp-Level Representations for Time Series with Hierarchical Contrastive Loss

**Authors:** Zhihan Yue, Yujing Wang, Juanyong Duan, Tianmeng Yang, Congrui Huang, Bixiong Xu  
**arXiv:** [2106.10466](https://arxiv.org/abs/2106.10466)  
**Date:** 2021-06-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2106.10466v1/

## Key points

- Authors Zhihan Yue, Yujing Wang, Juanyong Duan, Tianmeng Yang, Congrui Huang, and Bixiong Xu introduce TS2Vec - a framework for learning timestamp-level representations of time series data.
- TS2Vec utilizes timestamp-wise discrimination to generate contextual representation vectors for each timestamp directly.
- Learned representations exhibit superior predictive capabilities, with a linear regression model trained on top outperforming previous state-of-the-art techniques for supervised time series forecasting tasks.
- Instance-level representations can be easily derived by applying a max pooling layer over the learned representations of all timestamps.
- TS2Vec achieves significant performance improvements compared to existing methods for unsupervised time series representation across multiple datasets.
- The research findings are supported by comprehensive analysis and evaluation of the proposed framework's effectiveness in capturing temporal patterns within time series data.
- Availability of the source code on GitHub ensures reproducibility and further exploration by the research community.

## Summary

In their paper "Learning Timestamp-Level Representations for Time Series with Hierarchical Contrastive Loss," authors Zhihan Yue, Yujing Wang, Juanyong Duan, Tianmeng Yang, Congrui Huang, and Bixiong Xu introduce TS2Vec - a versatile framework for learning timestamp-level representations of time series data. Unlike traditional methods, TS2Vec utilizes timestamp-wise discrimination to directly generate contextual representation vectors for each timestamp in the time series. The study reveals that these learned representations exhibit superior predictive capabilities. In fact, a linear regression model trained on top of these representations outperforms previous state-of-the-art techniques for supervised time series forecasting tasks. Furthermore, the paper demonstrates that instance-level representations can be easily derived by applying a max pooling layer over the learned representations of all timestamps. Extensive experiments conducted on various time series classification tasks validate the quality of these instance-level representations. Notably, TS2Vec achieves significant performance improvements compared to existing state-of-the-art methods for unsupervised time series representation across 125 UCR datasets and 29 UEA datasets. The research findings are supported by a comprehensive analysis and evaluation of the proposed framework's effectiveness in capturing temporal patterns within time series data. The availability of the source code on GitHub (https://github.com/yuezhihan/ts2vec) ensures reproducibility and further exploration by the research community. Overall,"Learning Timestamp-Level Representations for Time Series with Hierarchical Contrastive Loss" presents a novel approach that advances the field of time series analysis and forecasting through innovative representation learning techniques.