Clustering of Time Series Regarding Their Over-Time Stability
Gerhard Klassen, Martha Tatusch, Stefan Conrad · 2020
The clustering of time series data is still a challenging task. There are different approaches which consider either multiple time series or a single one. While some interpret the whole sequence as one feature vector, others examine subsequences or extract relevant features first. Because of these various perspectives, very different statements result. In this paper, we present the clustering algorithm C(OTS)2 for multivariate time series data sets, that delivers a clustering per time point. It not only optimizes the quality of the clusters regarding intuitive demands, such as the spatial closeness of objects to their neighborhood within a cluster, but also the stability over time. Additionally, it can easily handle missing data points. The algorithm is of benefit whenever a cohesion of groups of time series can be assumed. One advantage is, that it requires only one parameter. Our experiments on different synthetic and real world data sets show, that our method works reasonable and fulfills the intention of finding temporal stable clusters without presupposing that the exact courses of the time series resemble.