Efficiency of Time Series Clustering Method Based on Distribution of Difference Using Several Distances

Abstract

Clustering is a machine learning method widely used in time series analysis. In this work, we cluster time series by applying four distance functions: Euclidean distance, Kullback-Leibler divergence, Wasserstein distance, and dynamic time warping. We consider the distribution of the first-order difference of time series and compare time series using such distributions under each of the four distances. Then, we model each time series as a vertex of a graph and the distance between each pair of time series as a weighted edge. Graph partitioning is performed as a clustering method. The advantages and drawbacks of each method are discussed. The experimental results show that Euclidean distance and Kullback-Leibler divergence perform better and more efficient clustering than the other two.