Time series clustering method with label propagation based on centrality

In order to cluster time series automatically and describe be the structural relations between of time series in more detail, this paper introduces the community discovery method to study time series clustering. According to the ability of the label propagation method which has limitation of uncertainty in the process and the sensitivity of the algorithm to the network structure, a clustering method for time series with label propagation based on centrality is proposed. Time series network structure is built, each time series is treat as a node in the network, and an updating order of labels is obtained according to each node's centrality. The membership degree of each node belonging to each community is computed, and the community is divided using belonging degree and label propagation, so as to achieve time series clustering. The proposed method analyzes the connection relationships among time series to find the structure features in the Euclidean space, thereby achieing the valid division of space structure. The experimental results demonstrate that the proposed clustering method does not need to determine the initial cluster center objects. It not only can detect simulated data network and real social network, but also obtains better results in time series clustering. © 2018, Editorial Office of Control and Decision. All right reserved.