Self-Supervised Community Detection Algorithm Based on Node Feature Convolution

Community detection is an important research topic in network science and has received extensive attention. Community detection aims to reveal clusters or group structures in complex networks, and many methods have been proposed, including those based on graph convolutional network (GCN). However, existing GCN model-based community detection methods rely excessively on prior labels. Moreover, GCN does not consider the difference in node degree distribution, resulting in poor node feature expression ability. To address the aforementioned issues, a self-supervised community detection algorithm based on node feature convolution (SGCN-NFC) is proposed. First, SGCN-NFC aggregates the neighbor nodes with high cosine similarity and construct a fixed-size feature map, according to node feature vectors. Then, a node feature convolution layer is introduced to assign different weights to each node to learn node features. Finally, a self-supervised module with semantical alignement is utilized to reduce the algorithm's dependence on known labels. Experimental results show that SGCN-NFC can bring better performance improvement in community detection compared with several state-of-the-art methods.