一种基于同步动力学模型的层次聚类方法

Abstract

Clustering is an essential method for analyzing and mining the intrinsic group in data. This paper presents a novel synchronization-based hierarchical clustering method based on an extended Kuramoto dynamic synchronization model. Each data object is regarded as a phase oscillator and interacts dynamically with its neighboring objects. As time evolves, objects synchronize naturally. With regard to the local diameter of the neighborhood, the proposed method nds local synchronization-based natural clusters. Hierarchical clustering results are achieved by enlarging the local neighborhood distance of objects synchronizing continuously. Using a neighborhood closure, our method predicts clusters before the objects reach local synchronization, thereby signi cantly reducing the dynamic interaction time. To select the optimal clusters automatically, this hierarchical clustering method based on a dynamic synchronization model is combined with a clustering validation measure known as the silhouette width criterion. Combined with the silhouette width criterion, the proposed method is parameter-free. Moreover, the proposed method can detect clusters in data of arbitrary shapes, sizes and numbers without any data distribution assumptions. This synchronization-based clustering also allows natural outlier identi cation, since outliers do not synchronize with data objects in clusters. Extensive experiments on several synthetic and real-world data sets demonstrate that the proposed method achieves high clustering accuracy with lower execution time and fewer synchronization steps compared to the state-of-the-art method.