An Adaptable Time Warping Distance for Time Series Learning

Most machine learning and data mining algorithms for time series datasets need a suitable distance measure. In addition to classic p-norm distance, numerous other distance measures exist and the most popular is dynamic time warping. Here we propose a new distance measure, called adaptable time warping (ATW), which generalizes all previous time warping distances. We present a learning process using a genetic algorithm that adapts ATW in a locally optimal way, according to the current classification issue we have to resolve. It's possible to prove that ATW with optimal parameters is at least equivalent or at best superior to the other time warping distances for all classification problems. We show this assertion by performing comparative tests on two real datasets. The originality of this work is that we propose a whole learning process directly based on the distance measure rather than on the time series themselves