Multivariate Time Series Representation and Similarity Search Using PCA

摘要

Multivariate time series (MTS) data mining has attracted much interest in recent years due to the increasing number of fields requiring the capability to manage and process large collections of MTS. In those frameworks, carrying out pattern recognition tasks such as similarity search, clustering or classification can be challenging due to the high dimensionality, noise, redundancy and feature correlated characteristics of the data. Dimensionality reduction is consequently often used as a preprocessing step to render the data more manageable. We propose in this paper a novel MTS similarity search approach that addresses these problems through dimensionality reduction and correlation analysis. An important contribution of the proposed technique is a representation allowing to transform the MTS with large number of variables to a univariate signal prior to seeking correlations within the set. The technique relies on unsupervised learning through Principal Component Analysis (PCA) to uncover and use, weights associated with the original input variables, in the univariate derivation. We conduct numerous experiments using various benchmark datasets to study the performance of the proposed technique. Compared to major existing techniques, our results indicate increased accuracy and efficiency. We also show that our technique yields improved similarity search accuracy.