With the development of science and technology, network scales of various fields have experienced an amazing growth. Networks in the fields of biology, economics and society contain rich hidden information of human beings in the form of connectivity structures. Network analysis is generally modeled as network partition and community detection problems. In this paper, we construct a community property-based triangle motif clustering scheme (COMICS) containing a series of high efficient graph partition procedures and triangle motif-based clustering techniques. In COMICS, four network cutting conditions are considered based on the network connectivity. We first divide the large-scale networks into many dense subgraphs under the cutting conditions before leveraging triangle motifs to refine and specify the partition results. To demonstrate the superiority of our method, we implement the experiments on three large-scale networks, including two co-authorship networks (the American Physical Society (APS) and the Microsoft Academic Graph (MAG)), and two social networks (Facebook and gemsec-Deezer networks). We then use two clustering metrics, compactness and separation, to illustrate the accuracy and runtime of clustering results. A case study is further carried out on APS and MAG data sets, in which we construct a connection between network structures and statistical data with triangle motifs. Results show that our method outperforms others in both runtime and accuracy, and the triangle motif structures can bridge network structures and statistical data in the academic collaboration area.
Keywords: Clustering; Community property; Large network; Triangle motif.
© 2019 Feng et al.