Change-point detection with deep learning: A review

Newswise — With the rapid growth of data complexity and volume, traditional statistical methods for change-point detection (CPD) face limitations in handling high-dimensional, non-stationary data and ensuring generalization across diverse scenarios. Deep learning techniques, however, have shown outstanding performance in time-series analysis tasks such as forecasting, enabling more robust and adaptive CPD methods. Despite the increasing adoption of deep learning in CPD, comprehensive literature reviews focusing on these modern approaches remain scarce, with previous surveys mostly centering on classical methods.

Therefore, Ruiyu XU, Zheren SONG, Jianguo WU, Chao WANG, and Shiyu ZHOU have jointly conducted a review titled “Change-point detection with deep learning: A review”. This review systematically explores the application of deep learning in CPD, aiming to fill the gap in existing literature. It first provides background information on CPD, including key definitions, the distinction between online and offline detection, and a detailed overview of its applications across various fields—such as healthcare (for seizure and arrhythmia detection), stock trading (for monitoring market crashes), industrial manufacturing (for predictive maintenance), speaker diarization, sleep research, and climate monitoring—along with the commonly used datasets in each domain. The review also elaborates on performance evaluation metrics for CPD algorithms, categorized into frame-based (e.g., Mean over Frames), CP-based (e.g., Mean Squared Error), and segment-based (e.g., Edit Score, F1 score) metrics.

Furthermore, the review organizes deep learning-based CPD methods into supervised and unsupervised frameworks. For supervised methods, it details network structures like Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), CNN-RNN hybrid models, Temporal Convolutional Networks (TCNs), Encoder-Decoder architectures, Autoencoders (AEs), and attention mechanisms, as well as relevant loss functions and their advantages (high accuracy with labeled data) and disadvantages (dependency on large labeled datasets, risk of overfitting). For unsupervised methods, it covers network structures including Multilayer Perceptrons (MLPs), AEs, RNNs, Graph Neural Networks (GNNs), self-supervised methods, and weakly-supervised methods, along with their loss functions and trade-offs (no need for labeled data but less interpretable results). The review also discusses preprocessing techniques (data transformation, denoising, data augmentation, windowing) and postprocessing methods (boundary refinement, continuity-based smoothing, system fusion, integration of prior knowledge) that enhance the effectiveness of CPD models.

Finally, the review identifies ongoing challenges in deep learning-based CPD, such as insufficient labeled data, difficulties in handling multi-modality data, demands for real-time online detection, limited model interpretability, and the need to integrate CPD results with decision-making. It also suggests potential future research directions to address these issues, including transfer learning, standardized multi-modal datasets, optimization of computational efficiency for online detection, and incorporation of domain-specific prior knowledge.

The paper “Change-point detection with deep learning: A review” is authored by Ruiyu XU, Zheren SONG, Jianguo WU, Chao WANG, Shiyu ZHOU. It is published in Front. Eng. Manag. 2025, 12(1): 154–176, with the DOI: open access at link.springer.com and journal.hep.com.cn.