Partition of Time Series in Temporal Axis Using Hierarchical Clustering

Abstract

Long-term time series data can be understood according to transitions of trends and features. We must partition time series into several terms of intervals of different trends and features. In statistics, change detection expresses it as a degree of change using probability distributions. In this paper, we propose a method to partition time series using a hierarchical clustering. First, we have clusters of one line segment connecting adjacent data. Then, we merge two adjacent similar clusters into one with a total similarity calculated by the weighted average of three similarities of values, change of values and oscillations. However, the partition results with the fixed weights do not fit our sense. We, therefore, propose variable weights with three similarities and sizes of adjacent clusters. Furthermore, in order to exclude small clusters including outliers, we define a similarity of two clusters adjacent to the small cluster. We apply this method to actual time series and show results.