Knowledge-Guidance Based Directed Graph Clustering.

Asymmetric relationships among objects are usually modeled by directed graphs, where each vertex has two types of edges, namely the edges connected from it and the edges connected to it. These two types of edges constitute two types of structural information, which is termed as out-structural information and instructural information. The aim of directed graph clustering is to assign vertexes into different clusters, such that the similarities between vertexes in the same cluster are higher than those from different clusters. Most of the existing algorithms for directed graph clustering define a similarity measure between vertexes by the density of edges between them. But in many cases, this kind of similarity measure is inappropriate. Thus, a novel definition of clusters for directed graphs is proposed in this paper, where the vertexes in the same cluster are more similar in terms of connecting patterns. After that, a novel Knowledge-Guidance based Directed Graph Clustering (KG-DGC) algorithm is proposed. In this method, the clustering result of one type of structural information is regarded as knowledge and fused in the clustering process of another type of structural information to get the overall clustering result. In order to evaluate the performance of the KG-DGC algorithm, two novel evaluation indices, estimating the similarity of connecting patterns between vertexes in the same cluster, are designed. The superior performance of the KG-DGC algorithm has been manifested by comparing with the state-of-art methods in the experiments.