Constructing High Dimensional Feature Space for Time Series Classification

摘要

The paper investigates a generic method of time series classification that is invariant to transformations of time axis. The state-of-art methods widely use Dynamic Time Warping (DTW) with One-Nearest-Neighbor (INN). We use DTW to transform time axis of each signal in order to decrease the Euclidean distance between signals from the same class. The predictive accuracy of an algorithm that learns from a heterogeneous set of features extracted from signals is analyzed. Feature selection is used to filter out irrelevant predictors and a serial ensemble of decision trees is used for classification. We simulate a dataset for providing a better insight into the algorithm. We also compare our method to DTW+1NN on several publicly available datasets.