A time series clustering method based on hypergraph partitioning

Abstract

Time series clustering has been applied in many scientific domains and has attracted much attention in recent years. In this paper, a novel hypergraph based clustering method for time series is proposed, combining multiple similarity measures and hypergraph partitioning. We firstly build the hypergraph for time series dataset using multiple similarity measures, where each time series is represented by vertex and hyperedge is formed based on the similarity relation among time series. Then, the vertices in the constructed hypergraph are grouped by the hypergraph partitioning method to identify the clusters of time series. Two different strategies for hypergraph construction are presented in detail, resulting in two specific methods. Empirical experiments of time series clustering with the UCR archive are conducted for the purpose of evaluation. The results demonstrate that the proposed methods outperform contrast clustering methods in most of the tested datasets and also achieve better performance than the network based methods under consideration, owe to the combination of advantages of hypergraph and various similarity measures.