Change Point Detection via Synthetic Signals

Abstract

Detecting change points in time series data is a widely acknowledged challenge with diverse applications, in which the data obtained from measured values is often characterized by complex compositions, and the availability of real data is typically limited. However, current detection algorithms often depend on domain-specific data to achieve better performance or are restricted to analyzing single variant series, limiting their applicability. In this paper, we introduce a novel approach to change point detection that eliminates the requirement for collecting supervised data. Initially, we train a discriminant model using artificially generated synthetic signals comprising a combination of intricate patterns and random noise. This discriminant model is designed to predict the number of change points, and the synthetic data set encompasses a wide range of patterns observed in real data and offers significant advantages in terms of diversity and data volume. The trained discriminant model is then applied in conjunction with the ClaSP method for change point detection. To fully exploit multivariate series information, we propose a simple yet useful weighted-merging method that improves detection performance by aggregating change point votes within each time gap. Experimental results demonstrate the superiority of our Detection Model via Synthetic Signals (DMSS) compared to the original ClaSP method, demonstrating exceptional performance on the Human Activity Segmentation dataset.