Metric learning in multiclass time series classification problem

This paper is devoted to the problem of multiclass time series classification. It is proposed to align time series in relation to class centroids. Building of centroids and alignment of time series is carried out by the dynamic time warping algorithm. The accuracy of classification depends significantly on the metric used to compute distances between time series. The distance metric learning approach is used to improve classification accuracy. The metric learning procedure modifies distances between objects to make objects from the same cluster closer and from the different clusters more distant. The distance between time series is measured by the Mahalanobis metric. The distance metric learning procedure finds the optimal transformation matrix for the Mahalanobis metric. To calculate quality of classification, a computational experiment on synthetic data and real data of human activity recognition was carried out.