Time series clustering based on relationship network and community detection

Abstract

Clustering is a fundamental part in data mining. Recent years have gone up in research on related fields owing to the widespread existence of time series in various fields. To describe the spatial relationship more accurately, our study led us to develop a novel clustering method based on complex network construction (nwNMF). The nwNMF includes a new method for mapping a univariate time series (UTS) dataset into a relationship network, where each object is represented by a corresponding network node, and similar nodes are connected by edges. The proposed solution is innovative based on the relationship network and relies on the use of community detection technology to achieve complete clustering, consisting of three stages. First, we recommend shape-based distance (SBD) to describe the similarity between two UTSs while improving time efficiency. Next, combining the idea of ϵ-nearest neighbor algorithm (ϵ-NN) to build the relationship network, in the process, we try to find the optimal parameter ϵ∗ and introduce a weight factor to explain the state of space more truthfully. Finally, community detection technology called matrix factorization to obtain a fuzzy membership matrix and generate clustering results. The proposed nwNMF is experimentally compared with other five classic clustering methods, and a simulation experiment is intended to illustrate the detailed network construction process. Experimental numerical results on various datasets show that nwNMF can improve the clustering accuracy and promote the clustering quality of UTS.