This repository contains a reading list of papers on Time Series Segmentation. This repository is still being continuously improved.
As a crucial time series preprocessing technique, semantic segmentation divides poorly understood time series into several discrete and homogeneous segments. This approach aims to uncover latent temporal evolution patterns, detect unexpected regularities and regimes, thereby rendering the analysis of massive time series data more manageable.
Time series segmentation often intertwines with research in many domains. Firstly, the relationship between time series segmentation, time series change point detection, and some aspects of time series anomaly/outlier detection is somewhat ambiguous. Therefore, this repository includes a selection of papers from these areas. Secondly, time series segmentation can be regarded as a process of information compression in time series, hence papers in this field often incorporate concepts from information theory (e.g., using minimum description length to guide the design of unsupervised time series segmentation models). Additionally, the task of decomposing human actions into a series of plausible motion primitives can be addressed through methods for segmenting sensor time series. Consequently, papers related to motion capture from the fields of computer vision and ubiquitous computing are also included in this collection.
Generally, the subjects of unsupervised semantic segmentation can be categorized into:
univariate time series: , where is the length of the time series.
multivariate time series: , where is the number of variables (channels).
tensor: , where denotes the dimensions other than time and variables.
In the field of time series research, unlike time series forecasting, anomaly detection, and classification/clustering, the number of papers on time series segmentation has been somewhat lukewarm in recent years (this observation may carry a degree of subjectivity from the author). Additionally, deep learning methods do not seem to dominate this area as they do in others. Some classic but solid algorithms remain highly competitive even today, with quite a few originating from the same research group. Therefore, in the following paper list, I will introduce them indexed by well-known researchers and research groups in this field.
Citation
🚩 2026/4/5: If you find this repository helpful for your research, please consider citing the following paper:
Zezhong Li, Ling Wang (2026). JT3D: Joint temporal segmentation for common pattern discovery in 3D tensor time series via segmented Gaussian modeling and information-gain. Neurocomputing, 682, 133453. https://doi.org/10.1016/j.neucom.2026.133453.
Zezhong Li, Wei Guo, Jianpeng An, Qi Wang, Yingchun Mei, Rongshun Juan, Tianshu Wang, Yang Li, and Zhongke Gao, AD2T: Multivariate Time Series Anomaly Detection with Association Discrepancy Dual-Decoder Transformer,IEEE Sensors Journal (IEEE Sens J), 2025. https://ieeexplore.ieee.org/document/10906418.
Ling Wang, Qinchun Xu, Zezhong Li, STANAP: Spatiotemporal Adaptive Neighborhood-Aware Prediction with Multimodal Evolution,Knowledge-Based Systems (KBS), 2026. https://doi.org/10.1016/j.knosys.2026.116346.
@article{LI2026JT3D,
title = {JT3D: Joint temporal segmentation for common pattern discovery in 3D tensor time series via segmented Gaussian modeling and information-gain},
author = {Zezhong Li and Ling Wang},
journal = {Neurocomputing},
volume = {682},
pages = {133453},
year = {2026},
issn = {0925-2312},
doi = {10.1016/j.neucom.2026.133453},
url = {https://www.sciencedirect.com/science/article/pii/S0925231226008507}
}
@ARTICLE{Li2025AD2T,
author={Li, Zezhong and Guo, Wei and An, Jianpeng and Wang, Qi and Mei, Yingchun and Juan, Rongshun and Wang, Tianshu and Li, Yang and Gao, Zhongke},
journal={IEEE Sensors Journal},
title={AD2T: Multivariate Time-Series Anomaly Detection With Association Discrepancy Dual-Decoder Transformer},
year={2025},
volume={25},
number={7},
pages={11710-11721},
keywords={Anomaly detection;Sensors;Transformers;Training;Time series analysis;Long short term memory;Correlation;Servers;Sensor phenomena and characterization;Measurement;Anomaly detection;association discrepancy;dual-decoder;multivariate time series (MTS);transformer},
doi={10.1109/JSEN.2025.3543835}}
@article{WANG2026STANAP,
title = {STANAP: Spatiotemporal adaptive neighborhood-aware prediction with multimodal evolution},
journal = {Knowledge-Based Systems},
volume = {348},
pages = {116346},
year = {2026},
issn = {0950-7051},
doi = {https://doi.org/10.1016/j.knosys.2026.116346},
url = {https://www.sciencedirect.com/science/article/pii/S0950705126010725},
author = {Ling Wang and Qinchun Xu and Zezhong Li},
}
Some Additional Information
🚩 2024/4/28: In fact, manually annotating segment points (change points) in large time series datasets is extremely labor-intensive and somewhat subjective. Therefore, the field of time series segmentation lacks large public datasets with ground truth, making it difficult for supervised methods to find sources of training data. Unsupervised time series segmentation also acts to some extent as an automatic annotator of segmentation points, making it easier to implement. Currently, 95% of the research work included in this repository is unsupervised.
🚩 2024/1/27: I have marked some recommended papers / datasets / implementations with 🌟 (Just my personal preference 😉).
Survey & Evaluation
NOTE: the ranking has no particular order.
TYPE
Venue
Paper Title and Paper Interpretation
Code
Survey
Front. Eng. Manag. '25
Change-point detection with deep learning: A review
None
Survey
J R STAT SOC B. (JRSSB) '24
Automatic change-point detection in time series via deep learning
univariate time series:
, where
is the length of the time series.
multivariate time series:
, where
is the number of variables (channels).
tensor:
, where
denotes the dimensions other than time and variables.