PCA-based Time Series Similarity Search

Abstract

We propose a novel approach in multivariate time series similarity search for the purpose of improving the efficiency of data mining techniques without substantially affecting the quality of the obtained results. Our approach includes a representation based on principal component analysis (PCA) in order to reduce the intrinsically high dimensionality of time series and utilizes as a distance measure a variation of the squared prediction error (SPE), a well-known statistic in the Statistical Process Control community. Contrary to other PCA-based measures proposed in the literature, the proposed measure does not require applying the computationally expensive PCA technique on the query. In this chapter, we investigate the usefulness of our approach in the context of query by content and 1-NN classification. More specifically, we consider the case where there are frequently arriving objects that need to be matched with the most similar objects in a database or that need to be classified into one of several pre-determined classes. We conduct experiments on four data sets used extensively in the literature, and we provide the results of the performance of our measure and other PCA-based measures with respect to classi- fication accuracy and precision/recall. Experiments indicate that our approach is at least comparable to other PCA-based measures and a promising option for similarity search within the data mining context.