Time series reconstruction analysis

The dimensionality of time series data is usually very large, so it must often be reduced before applying certain data mining tasks upon it. Dimensionality reduction is achieved by creating appropriate time series representation that is actually new time series of lower dimensionality obtained from the original one by preserving only the important features. I addition to that, the reconstruction of original time series from its representation is inevitable task in many practical applications. The main objective of this paper is the comparison of different time series representations in the term of their reconstruction accuracies on a number of freely available data sets. Reconstructed time series are compared with the original ones, using several state-of-the-art similarity measures, in order to measure the quantity of information loss. Additionally, we measured the correlations between several data set properties and their reconstruction errors, which will give a deeper insight in the problem of choosing of appropriate representation technique for a particular data set.