Comparing three time series segmentation methods via novel evaluation criteria

Abstract

Time series segmentation is one of the important basic tasks in time series data mining. Indeed, it is a pre-requisite step for many other time series mining tasks such as dimension reduction, representation, classification, clustering, prediction, motif discovery and anomaly detection in time series. Since several segmentation techniques have been proposed, it is difficult to select a suitable one for a specific application if we do not know how to evaluate the performance of various segmentation methods. This paper focuses on comparing the quality of three well-known segmentation methods: Important Extreme Points, Perceptually Important Points and Polynomial Least Square Approximation by using some existing evaluation criteria for time series segmentation. However, the existing evaluation criteria required prior knowledge about the time series and their segments as external information in their criteria. Therefore, we propose one more evaluation criterion for time series segmentation. Experimental results have showed that the set of novel evaluation criteria can help to quantify segmentation results and the quality of each segmentation method depends on the characteristics of each tested dataset.