Effect of Mahalanobis Distance on Time Series Classification Using Shapelets

Abstract

The sequence of values that are measured at time intervals equally spaced is time series data. Finding shapelets within a data set as well as classifying that data based on shapelets is one of the most recent approaches to classification of this data. In the classification using shapelets, Euclidean distance measure is adopted to find dissimilarity between two time series sequences. Though the Euclidean distance measure is known for its simplicity in computation, it has some disadvantages: it requires data to be standardized and it also requires that the two data objects being compared be of the same length. It is sensitive to noise as well. To overcome the problem, Mahalanobis distance measure can be used. In the proposed work, classification of time series data is performed using time series shapelets and used Mahalanobis distance measure which is the measure of distribution between a point and distribution. Correlations between data set is considered. It does not depend on scale. The cost complexity pruning is performed on decision tree classifier. The Mahalanobis distance improves the accuracy of algorithm and cost complexity pruning method reduces the time complexity of testing and classification of unseen data. The experimental results show that the Mahalanobis distance measure leads to more accuracy and due to decision tree pruning the algorithm is faster than existing method.