DTW as Alignment Function in the Context of Time Series Balancing

Abstract

Frequently, big data problems deal with Time Series (TS) datasets. This identification of positive events (falls, epilepsy crisis, stroke crisis, etc.) in this kind of problems needs a suitable number of the positive class TS. Consequently, the machine learning algorithms need to cope with the TS data balancing problem, which has not been studied in great depth in the literature. In one of our previous works we presented a TS extension (TS_SMOTE) of the well-known SMOTE algorithm for balancing datasets. The specification of this new algorithm consider two TS distance functions: (i) one used (as KNN distance measure) to choose the parents TS for each new synthetic TS, and (ii) a second distance function used as alignment function of the parents to obtain the new synthetic TS. As there are plenty of alignment functions in the literature here we have chosen the well-known DTW distance and compared against the EUCLIDEAN distance. Thus, current work presents an analysis of different parameters involved in our TS_SMOTE algorithm in order to prove the validity of DTW as good alignment measurement in the context of TS balancing. Some of the parameters analysed are: (a) criteria to choose the parents for new TS in the KNN algorithm, (b) the k parameter of KNN and finally (c) how the level of un alignment between the TS in the dataset affects the TS_SMOTE results. We can conclude that DTW has a very good behaviour obtaining new balanced datasets when the original TS are aligned but not quite good when the TS are unaligned. Also, the EUCLIDEAN distance get very poor results in all the experiments.