An analysis of the relevance of temporal information in time series classification

Abstract

In recent years, the interest in time series has increased considerably due to the vast amount of such data collected in a variety of fields. Time series are a particular type of data: they are sets of ordered observations. The analysis of databases composed of time series requires the consideration of the nature of the instances, where there exist a temporal correlation among the observations. A considerable variety of algorithms that take heed of such characteristic of the instances have been developed to represent, index, cluster and classify time series. Particularly, this work focuses on the classification task. Firstly, a review is performed about the specific time series classifiers, to give an insight of their workflow as well as how they capture the temporal information. The different procedures of those classifiers endow them with different abilities to catch the intrinsic temporal information of the instances for classification. Moreover, this work carries out an experiment based on empirical distributions to estimate the sensitivity of specific time series classifiers to the temporal order of the observations. Besides, although in general specific classifiers have been used to classify time series, in some time series classification problems, non-specific classification algorithms have shown to be competitive with the specific ones. Thus the relevance of the temporal order for classification varies for different time series classification problems. The present work aims to develop an analysis based on empirical distributions for estimating the relevance of the temporal ordering in a given time series classification problem, as well as studying the sensitivity of the specific time series classifiers to the temporal correlation of the observations.

English