An angle-based dissimilarity for accelerating the clustering of dynamic data in networks

Mining time series data is of great significance in various areas. In order to efficiently find representative patterns in these data sets, this paper focuses on the definition of a valid dissimilarity and the acceleration of partitioning clustering, a common technique used to discover typical shapes of time series. Following the analysis of adopting the angle between two time series as the measure of dissimilarity, our definition, which is invariant to specific transformations, has been proposed. Moreover, our dissimilarity obeys the triangle inequality with specific restrictions. This property can be employed to accelerate clustering. An integrated algorithm is proposed. Experiments show that the angle-based dissimilarity captures the essence of time series patterns that are invariant to amplitude scaling. In addition, our algorithm provides a feasible way to update cluster centers, as well as an effective approach to accelerating clustering. Our accelerated algorithm reduces the number of dissimilarity calculations by almost an order of magnitude.