DPSCAN: Structural Graph Clustering Based on Density Peaks

Structural graph clustering is one of the fundamental problems in managing and analyzing graph data. The structural clustering algorithm SCAN is successfully used in many applications because it obtains not only clusters but also hubs and outliers. However, the results of SCAN heavily depend on two sensitive parameters, epsilon and mu, which may result in loss of accuracy and efficiency. In this paper, we propose a novel Density Peak-based Structural Clustering Algorithm for Networks (DPSCAN). Specifically, DPSCAN clusters vertices based on the structural similarity and the structural dependency between vertices and their neighbors, without tuning parameters. Through theoretical analysis, we prove that DPSCAN can detect meaningful clusters, hubs and outliers. In addition, to improve the efficiency of DPSCAN, we propose a new index structure named DP-Index, so that each vertex needs to be visited only once. Finally, we conduct comprehensive experimental studies on real and synthetic graphs, which demonstrate that our new approach outperforms the state-of-the-art approaches.