A Study on Ward Method for Time-Series Data

Abstract

Time-series data is data in which values are given at regular intervals over time. Clustering is one of the analysis methods for time-series data. There are two dissimilarity measures for time-series data: Dynamic Time Warping (DTW) and Shape-based distance (SBD). Hierarchical clustering is a method to visualize the process of cluster partition as a dendrogram. It is known that the Ward method assumes Euclidean space, and it is usually inappropriate to apply dissimilarities such as DTW or SBD. In this paper, we consider the way to apply the dissimilarity matrices obtained from DTW and SBD as a kernel function to the Ward method. The proposed method takes the same approach as previous studies on the Ward method. Numerical experiments were conducted on six benchmark datasets and one artificial dataset. The proposed method tends to obtain better cluster partitions than existing clustering methods.