Density-based Clustering of Time Series Subsequences

Doubts have been raised that time series subsequences can be clustered in a meaningful way. This paper introduces a kernel-density-based algorithm that detects meaningful patterns in the presence of a vast number of random-walk-like subsequences. The value of density-based algorithms for noise elimination in general has long been demonstrated. The challenge of applying such techniques to time-series data consists in first specifying uninteresting sequences that are to be considered as noise, and secondly ensuring that those uninteresting sequences will not affect the clustering result. Both problems are addressed in this paper and the success of the technique is demonstrated on several standard data sets.