Community detection based on community perspective and graph convolutional network

Abstract

Community detection is an essential topic in network analysis, which aims to divide a network into multiple subgraphs to mine potential information. However, most existing approaches tend to separate representation learning from clustering and fail to detect overlapping communities. Furthermore, these methods do not take a community perspective and cannot effectively capture information at the community level. In this paper, we propose a new community detection method based on the community perspective and graph convolution network (CPGC) to address these limitations in attributed networks without prior label information. First, through the Bernoulli-Poisson model, CPGC combines representation learning and clustering and can be used for overlapping communities. Second, we modify the classical graph convolution to enhance the discriminability of node representations, making them more suitable for increasingly large network data. Finally, we propose a novel community perspective similarity and introduce cross-community modularity to leverage community-level information. These improvements enable CPGC to be community-oriented and explore potential community structures more accurately. Experiments on various real-world networks show that CPGC can achieve state-of-the-art results in nonoverlapping or overlapping communities.