An Enhanced Parameter-Free Subsequence Time Series Clustering for High-Variability-Width Data

Abstract

In time series mining, subsequence time series (STS) clustering has been widely used as a subroutine in various mining tasks, e.g., anomaly detection, classification, or rule discovery. STS clustering’s main objective is to cluster similar underlying subsequences together. Other than the known problem of meaninglessness in the STS clustering results, another challenge is on clustering where the subsequence patterns have variable lengths. General approaches provide a solution only to the problems where the range of width variability is small and under some predefined parameters, which turns out to be impractical for real-world data. Thus, we propose a new algorithm that can handle much larger variability in the pattern widths, while providing the parameter-free characteristic, so that the users would no longer suffer from the difficult task of parameter selection. The Minimum Description Length (MDL) principle and motif discovery technique are adopted to be used in determining the proper widths of the subsequences. The experimental results confirm that our proposed algorithm can effectively handle very large width variability of the time series subsequence patterns by outperforming all other recent STS clustering algorithms.