Deep graph clustering with enhanced feature representations for community detection

Community detection is to partition community nodes into groups with similar attributes and topologies. In recent years, community detection becomes a research hotspot due to its great value and broad applications to social sciences. Thus, many clustering methods are developed for community detection. In particular, deep graph clustering has become a mainstream community detection approach because of its powerful abilities of feature representation and relationship extraction. Deep graph clustering uses graph neural networks (e.g., graph autoencoders) to learn the node feature representations with abundant topological relationships. Though deep graph clustering succeeds in dealing with topological relationships of nodes, there are insufficient attribute information in its learned feature representations, which hurt detection performance. To solve this problem, we propose an enhanced feature representation approach for deep graph clustering in community detection. Firstly, we construct a basic autoencoder to learn hierarchical attribute information and then deliver it into neural layers of a graph autoencoder. The graph autoencoder organically combines the received hierarchical attribute information with its extracted topological relationships to generate enhanced feature representations for clustering. Secondly, we design a self-supervised mechanism to optimize our deep graph clustering model. This mechanism uses the reconstruction losses of two autoencoders and clustering loss as self-supervised information to efficiently guide model updates. In this way, our approach overcomes the insufficient attribute information in generated feature representations, thus is more conducive to community detection. Extensive experiments demonstrate that the deep graph clustering with enhanced feature representations improves the performance of community detection compared to the other popular deep graph clustering approaches.