A Statistic-Augmented Neural Network for change point detection and type identification

Change point detection and type classification remain challenging across disciplines. Classical procedures rely on strict distributional assumptions, while deep learning demands large training sets and high computational cost. To address these limitations, this article proposes a Statistic-Augmented Neural Network (SNN) that unifies both paradigms. The method employs the Cumulative Sum (CUSUM) statistic as an augmentation module, whose output feeds into a lightweight neural network. This hybrid design enables simultaneous detection and type identification without prior assumptions on distributions or change point modalities, achieving robust inference even under limited samples. Ablation studies confirm that CUSUM augmentation amplifies signal representation, reducing reliance on large-scale training data. Extensive simulations assess performance across diverse scenarios, and the method is applied to human activity recognition from wearable motion sensors using the Human Activity Sensing Consortium (HASC) dataset. The framework offers a practical solution for detecting distributional shifts in sequential data across a broad range of applications.