Reproduce “Automatic Change-Point Detection in Time Series via Deep Learning” on OpenResearch

About this paper

Jie Li, Paul Fearnhead, Piotr Fryzlewicz, Tengyao Wang

The authors train neural networks to detect changes in time series by classifying labeled examples as change or no change, learning a decision rule rather than tuning an existing test. A network can represent CUSUM for mean changes. In simulations, networks matched tuned CUSUM under independent Gaussian noise and had lower test misclassification error under tested autocorrelated and heavy-tailed noise.

Key insights

• Classical statistical change-point tests, such as the CUSUM test and generalized likelihood-ratio tests, can be explicitly represented as simple neural networks, providing a theoretical foundation for the deep learning approach.
• Neural networks can learn robust detection features directly from data, enabling them to significantly outperform traditional methods when underlying model assumptions (e.g., Gaussianity, independence) are violated.
• By leveraging labelled examples for supervised learning, the framework automates the test statistic construction and parameter tuning, simplifying the change-point detection workflow for practitioners.

Summary by alphaXiv.