Applying Prototype Selection and Abstraction Algorithms for Efficient Time-Series Classification

A widely used time series classification method is the single nearest neighbour. It has been adopted in many time series classification systems because of its simplicity and effectiveness. However, the efficiency of the classification process depends on the size of the training set as well as on data dimensionality. Although many speed-up methods for fast time series classification have been proposed and are available in the literature, state-of-the-art, non-parametric prototype selection and abstraction data reduction techniques have not been exploited on time series data. In this work, we present an experimental study where known prototype selection and abstraction algorithms are evaluated both on original data and a dimensionally reduced representation form of the same data from seven popular time series datasets. The experimental results demonstrate that prototype selection and abstraction algorithms, even when applied on dimensionally reduced data, can effectively reduce the computational cost of the classification process and the storage requirements for the training data, and, in some cases, improve classification accuracy.