The Consideration of Proper Time-series Clustering

Abstract

A key issue for the clustering of time-series dataset is not being aware of the appropriate number of clusters. In particular, the true cluster of most time-series is not known. There are many interesting properties of time-series data such as scale, shape, structure, model and others. These features can make the different results in clustering of time-series. Two basic techniques used to determine the appropriate number of clusters in general are the silhouette and the sum of squared error, which are not enough for considering clusters of time-series data. In this research, we present a method for evaluating the suitability of cluster by using the silhouette and the sum of squared error. Furthermore, we integrate evaluation method with the designated representation using agent of time-series components. The experimentation uses both synthetic and 2 sets of real time-series data to compare the results of partitioning clustering and hierarchical clustering with 6 similarity measures. The results of our proposed evaluation method showed that the number of clusters was in accordance with the silhouette and the sum of squared error. Moreover the similarity was meaningful