Time Series Classification with Representation Ensembles

Abstract

Time series has attracted much attention in recent years, with thousands of methods for diverse tasks such as classification, clustering, prediction, and anomaly detection. Among all these tasks, classification is likely the most prominent task, accounting for most of the applications and attention from the research community. However, in spite of the huge number of methods available, there is a significant body of empirical evidence indicating that the 1-nearest neighbor algorithm ($1$-NN) in the time domain is “extremely difficult to beat”. In this paper, we evaluate the use of different data representations in time series classification. Our work is motivated by methods used in related areas such as signal processing and music retrieval. In these areas, a change of representation frequently reveals features that are not apparent in the original data representation. Our approach consists of using different representations such as frequency, wavelets, and autocorrelation to transform the time series into alternative decision spaces. A classifier is then used to provide a classification for each test time series in the alternative domain. We investigate how features provided in different domains can help in time series classification. We also experiment with different ensembles to investigate if the data representations are a good source of diversity for time series classification. Our extensive experimental evaluation approaches the issue of combining sets of representations and ensemble strategies, resulting in over 300 ensemble configurations.

This work partially funded by grant #2012/08923-8, #2013/26151-5, and #2015/07628-0, São Paulo Research Foundation (FAPESP); and CNPq #446330/2014-0 and #303083/2013-1.