Multivariate Time Series Clustering via Multi-relational Community Detection in Networks

Abstract

Clustering multivariate time series is a challenging problem with numerous applications. The presence of complex relations amongst individual series poses difficulties with respect to traditional modelling, computation and statistical theory. In this paper, we propose a method for clustering multivariate time series by using multi-relational community detection in complex networks. Firstly, a set of multivariate time series is transformed into a multi-relational network. Then, an algorithm for multi-relational community detection based on multiple nonnegative matrices factorization (MNMF) is proposed and is applied to identify time series clusters. The transformation of time series from time-space domain to topological domain benefits from the ability of networks to characterize both local and global relationship amongst nodes (representing data samples), while the use of MNMF can give full play to complex relations amongst individual series and preserve the multi-way nature of multivariate information. Preliminary experiment indicates promising results of our proposed approach.