Semi-Supervised Change-Point Detection via Consistency Training on Augmented Temporal Views

Abstract

The detection of abrupt changes in the generative parameters of time series data, known as Change-Point Detection (CPD), is a fundamental challenge in signal processing, data mining, and machine learning. While supervised deep learning methods have achieved remarkable success in this domain, they rely heavily on large-scale, accurately annotated datasets. In many real-world applications, such as industrial anomaly detection and physiological signal monitoring, obtaining frame-level labels is prohibitively expensive and requires expert domain knowledge. Conversely, unsupervised methods often suffer from high false-positive rates due to their inability to distinguish between noise and varying semantic states. To bridge this gap, this paper proposes a novel Semi-Supervised Change-Point Detection framework based on Consistency Training. We introduce a dual-branch architecture that enforces predictive consistency between original temporal sequences and their stochastically augmented views. By leveraging a rigorous set of temporal augmentations—including magnitude scaling, permutation, and time-warping—we enable the model to learn robust feature representations from unlabeled data while refining decision boundaries using a limited set of labeled examples. Extensive experiments on both synthetic and real-world datasets demonstrate that our approach significantly outperforms state-of-the-art unsupervised methods and achieves competitive performance against fully supervised baselines using only 10% of the labeled data.