Selecting Time Series Clustering Methods based on Run-Time Costs

Abstract

Clustering time series, e.g., of monitoring data from software systems, can reveal important insights and interesting hidden patterns. However, choosing the right method is not always straightforward, especially as not only clustering quality but also run-time costs must be considered. In this paper, we thus present an approach that aids users in selecting the best methods in terms of quality as well as computational costs. Given a set of candidate methods, we evaluate their clustering performance and robustly measure their actual run times, i.e., the execution time on a specific machine. We evaluate our approach using data from the UCR time series archive and show its usefulness in determining the best clustering methods while also taking costs into account.