Mining Massive Time Series Data: With Dimensionality Reduction Techniques

Abstract

A pre-processing step to reduce the volume of data but suffer an acceptable loss of data quality before applying data mining algorithms on time series data is needed to decrease the input data size. Input size reduction is an important step in optimizing time series processing, e.g. in data mining computations. During the last two decades various time series dimensionality reduction techniques have been proposed. However no study has been dedicated to gauge these time series dimensionality reduction techniques in terms of their effectiveness of producing a reduced representation of the input time series that when applied to various data mining algorithms produces good quality results. In this paper empirical evidence is given by comparing three reduction techniques on various data sets and applying their output to four different data mining algorithms. The results show that it is sometimes feasible to use these techniques instead of using the original time series data. The comparison is evaluated by running data mining methods over the original and reduced sets of data. It is shown that one dimensionality reduction technique managed to generate results of over 83% average accuracy when compared to its benchmark results.