A time series clustering technique based on community detection in networks

Abstract

Time series clustering is a research topic of practical importance in temporal data mining. The goal is to identify groups of similar time series in a data base. In this paper, we propose a technique for time series clustering via community detection in complex networks. First, we construct a network where every vertex represents a time series connected its most similar ones,. Similarity was calculated using different time series distance functions. Then, we applied a community detection algorithm to identify groups of strongly connected vertices in order to produce time series clusters. We verified which distance function works better with every clustering algorithm and compared them to our approach. The experimental results show that our approach statistically outperformed many traditional clustering algorithms. We find that the community detection approach can detect groups that other techniques fail to identify.