Shape-Based Conditional Neural Field for Wrist-Worn Change-Point Detection

Yuang Shi, Varsha Suresh, Wei Tsang Ooi

Abstract: We consider the problem of online change-point detection in a time-series, with application to detecting switching of human activities through accelerometer readings on a wrist-worn device. Traditional approaches for change-point detection do not work well on human activities in the wild and are too computationally expensive for implementation on wrist-worn devices. To address these shortcomings, we propose ShapeCNF, a simple and fast online change-point detection method that uses shape-based features to model the activity patterns and a conditional neural field to model the temporal correlations among the time series segments. Extensive experiments conducted on a highly dynamic time-series dataset demonstrate the effectiveness and superiority of ShapeCNF over the state-of-the-art methods.