Community-Oriented Graph Convolutional Network For Unsupervised Community Detection

Community detection is a task to partition a network into multiple communities,which has important theoretical and practical significance.Graph convolutional network(GCN),a new deep learning technique,has recently been developed for community detection.MRFas GCN(Markov Random Field as Graph Convolutional Networks),a new GCN method for community detection,goes one step further to incorporate Markov Random Fields(MRF)modeling of community structures into GCN model to further improve community detection performance.However,the existing GCN models are semi-supervised methods,even though community detection is in essence an unsupervised learning problem(this is mainly because little training data are available for most community applications and information from one network can be hardly used for another network).To address above problem,this paper introduced two GCN approaches for unsupervised community detection in a progressive way under the framework of Autoencoder.1)This paper first cast MRFas GCN model as an encoder to derive node community membership in the hidden layer of the encoder,and used the inner product of network embedding as a decoder to reconstruct the network topologies,so as to create a simple unsupervised community detection model.To validate the effectiveness of this method,this paper compared it with six representative community detection methods on nine real networks of different sizes.But experiments showed that this method had no better performance.2)In order to solve this problem,this paper further found through further experimental analysis that the decoder mechanism used in the above model was not suitable for community detection task and did not reconstruct both network topology and node attribute information at the same time.However,the existing methods have shown that it is generally more effective to use both topology and attribute information than to use one alone.Therefore,this paper introduced two kinds of information into the final unsupervised community detection GCN method.Specifically,this paper used the same encoder mechanism as the original model and then introduced a community-centric dual decoder to reconstruct network topologies and node attributes separately in an unsupervised fashion,for faithful community detection in the hidden space.This paper specifically introduced a scheme of local structural enhancement to accommodate nodes to have more common neighbors and similar attributes with the similar community memberships.Experimental results on some real-world networks showed that this paper proposed new method outperformed the representative methods for community detection.We also showed the effectiveness of the novel decoding mechanism for generating links and attributes together over the commonly used methods for reconstructing links alone.