Novel Criteria to Measure Performance of Time Series Segmentation Techniques

An important task in signal processing and temporal data mining is time series segmentation. In order to perform tasks such as time series classification, anomaly detection in time series, motif detection, or time series forecasting, segmentation is often a pre-requisite. However, there has not been much research on evaluation of time series segmentation techniques. The quality of segmentation techniques is mostly measured indirectly using the least-squares error that an approximation algorithm makes when reconstructing the segments of a time series given by segmentation. In this article, we propose a novel evaluation paradigm, measuring the occurrence of segmentation points directly. The measures we introduce help to determine and compare the quality of segmentation algorithms better, especially in areas such as finding perceptually important points (PIP) and other user-specified points.