Instance reduction for time series classification by exploiting representative characteristics using k-means

Abstract

1-Nearest Neighbor has been endorsed an efficient method for time series classification as it outperforms more advanced classification algorithms in most cases. However, the time and space efficiency of this method depend on the number of instances in a training set. In order to improve its running time and space using, one can apply an approach called instance reduction. This approach reduces the training set size by choosing the best instances and using only them during classification of new instances. However, the high-dimensional characteristic of time series is a challenge for many mining tasks including instance reduction. Due to this challenge, only two instance reduction methods have been proposed: Naïve Rank Reduction and INSIGHT. In this work, we propose a new approach to instance reduction using K-means. Our method can select a reasonable set of instances so that it can preserve the representative characteristics and keep the generalization of the original training set. We conduct experiments to evaluate our method on 46 public datasets from the UCR Classification Archive. The experimental results show that the proposed method achieves state-of-the-art performance on these datasets.