A novel clustering-based 1-NN classification of time series based on MDL principle

Abstract

In this work, we propose a clustering-based k-NN classifier for time series data. The classifier aims to select useful instances for the training set at the classifying step in order to reduce the training set and speed up the classification. First, we apply the MDL principle in selecting the core and peripheral sets for the clusters formed in the clustering step. Second, our classifier applies the Compression Rate Distance, a powerful distance measure for time series, which was proposed in our previous work. We conducted experiments over a vast majority number of time series datasets. The experimental results reveal that our proposed method can outperform the original Clustering-based k-NN and the two other methods INSIGHT, and Naïve Rank in most of the tested datasets. In comparison to the traditional k-NN method which use the whole original training set, our proposed method can run much faster while the accuracy rates decrease insignificantly (about 1.59 %) and in some data sets, the accuracy rates even increase.