Structured Loss for Deep Change-Point Detection

Change-point detection, which is finding abrupt changes in the behaviour of a signal, is a crucial step in numerous data processing pipelines. There are many available algorithms in the literature; choosing one and calibrating it is complex and time-consuming. In many applications, labels are available, but very few articles have proposed using that information to facilitate calibration. This work describes a methodology for learning a signal representation where change points are easily detected. Our approach relies on a loss function tailored for the change detection task, which we combine with a deep neural network to transform the signal. Experimental results on several real-world data sets show that our method improves segmentation accuracy over baseline methods.