Self-Supervised Learning for Time Series Analysis - Survey Paper

Oct 21, 2026

Uploaded by Eliana

Self-supervised learning for time series analysis reduces reliance on labeled data by using a pre-training and fine-tuning strategy. Compared with existing self-supervised surveys in computer vision and natural language processing, a comprehensive time-series SSL taxonomy is still lacking. This survey reviews state-of-the-art SSL methods for time series, organizes them into generative-, contrastive-, and adversarial-based categories, and further details subcategories, key ideas, frameworks, advantages and limitations.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages20

Reading Time50 min

AccessFree to download

Introduction

Time series tasks and feature extraction

Self-supervised learning as label-efficient learning

Self-Supervised Learning for Time Series Analysis: Taxonomy, Progress, and Prospects Kexin Zhang , Graduate Student Member, IEEE, Qingsong Wen, Senior Member, IEEE, Chaoli Zhang , Rongyao Cai , Graduate Student Member, IEEE, Ming Jin, Yong Liu , Member, IEEE, James Y. Zhang , Yuxuan Liang , Member, IEEE, Guansong Pang, Member, IEEE, Dongjin Song, Member, IEEE, and Shirui Pan , Senior Member, IEEE (Survey Paper) Abstract—Self-supervised learning (SSL) has recently achieved impressive performance on various time series tasks. The most prominent advantage of SSL is that it reduces the dependence on labeled data. Based on the pre-training and ﬁne-tuning strategy, even a small amount of labeled data can achieve high performance. Compared with many published self-supervised surveys on computer vision and natural language processing, a comprehensive survey for time series SSL is still missing. To ﬁll this gap, wereview current state-of-the-art SSL methods for time series data in this article. To this end, we ﬁrst comprehensively review existing surveys related to SSL and time series, and then provide a new taxonomy of existing time series SSL methods by summarizing them from three perspectives: generative-based, contrastive-based, and adversarial-based. These methods are further divided into ten Manuscript received 17 July 2023; revised 23 January 2024; accepted 6 April 2024 . Date of publication 10 April 2024; date of current version 5 September 2024 . This work was supported in part by the Key R&D Project of Zhejiang Province under Grant 2024C01172 and in part by the National Key R&D Program of China under Grant 2021YFB2012300 . Recommended for acceptance by M. Cho. (Corresponding authors: Qingsong Wen; Yong Liu.) Kexin Zhang is with the Huzhou Institute of Zhejiang University, Huzhou 313000, China, and also with the College of Control Science and Engineering, Zhejiang University, Hangzhou 310027, China (e-mail: [zhangkexin@zju.edu.cn](zhangkexin@zju.edu.cn)). QingsongWen was with Alibaba Group. Heisnow with Squirrel AI, Bellevue, WA 98004 USA (e-mail: [qingsongedu@gmail.com](qingsongedu@gmail.com)). Chaoli Zhang is with the School of Computer Science and Technology, Zhejiang Normal University, Jinhua, Zhejiang 321017, China (e-mail: [chaolizcl@zjnu.edu.cn](chaolizcl@zjnu.edu.cn)). Rongyao Cai and Yong Liu are with the College of Control Science and Engineering, Zhejiang University, Hangzhou 310027, China (e-mail: [rycai@zju.edu.cn](rycai@zju.edu.cn); [yongliu@iipc.zju.edu.cn](yongliu@iipc.zju.edu.cn)). Ming Jin is with the Faculty of Information Technology, Monash University, Clayton, VIC 3800, Australia ([e-mail: ming.jin@monash.edu](e-mail: ming.jin@monash.edu)). James Y. Zhang is with Ant Group, Hangzhou, Zhejiang 310058, China (e-mail: [james.z@antgroup.com](james.z@antgroup.com)). Yuxuan Liang is with the INTR & DSA Thrust, Hong Kong University of Science and Technology, Clear Water Bay, Hong Kong (e-mail: [yuxliang@outlook.com](yuxliang@outlook.com)). Guansong Pang is with the School of Computing and Information Systems, Singapore Management University, Singapore 188065 (e-mail: [gspang@smu.edu.sg](gspang@smu.edu.sg)). Dongjin Song is with the School of Computing, University of Connecticut, Storrs, CT 06269 USA (e-mail: [dongjin.song@uconn.edu](dongjin.song@uconn.edu)). Shirui Pan is with the School of Information and Communication Technology, Grifﬁth University, Southport, Qld 4222, Australia (e-mail: s.pan@grifﬁ[th.edu.au](th.edu.au)). [https://github.com/qingsongedu/Awesome-SSL4TS](https://github.com/qingsongedu/Awesome-SSL4TS). This article has supplementary downloadable material available at [https://doi.org/10.1109/TPAMI.2024.3387317](https://doi.org/10.1109/TPAMI.2024.3387317), provided by the authors. Digital Object Identiﬁer 10.1109/TPAMI.2024.3387317 subcategories with detailed reviews and discussions about their key intuitions, main frameworks, advantages and disadvantages.

FAQs

Why is self-supervised learning important for time series analysis?

It significantly reduces dependence on labeled data and improves performance via a pre-training and fine-tuning strategy, making it effective when labels are costly to obtain.

How does the survey categorize time series self-supervised learning methods?

It introduces a taxonomy that groups methods into generative-based, contrastive-based, and adversarial-based approaches, with further subcategories and discussions.

What resources does the survey include to support experiments and evaluation?

It summarizes datasets commonly used for time series forecasting, classification, anomaly detection, and clustering tasks to facilitate experimentation and validation.

Oct 21, 2026

Uploaded by Eliana