Clustering time series is usually limited by the fact that the length of the time series has a significantly negative influence on the runtime. On the other hand, approximative clustering applied to existing compressed representations of time series (e.g. obtained through dimensionality reduction) usually suffers from low accuracy. We propose a method for the compression of time series based on mathematical models that explore dependencies between different time series. In particular, each time series is represented by a combination of a set of specific reference time series. The cost of this representation depend only on the number of reference time series rather than on the length of the time series. We show that using only a small number of reference time series yields a rather accurate representation while reducing the storage cost and runtime of clustering algorithms significantly. Our experiments illustrate that these representations can be used to produce an approximate clustering with high accuracy and considerably reduced runtime.
展开▼
