Overlapping community detection based on graph attention autoencoder and self-trained clustering

Abstract

Existing methods for detecting overlapping communities often rely solely on the attributes of the nodes and the network structure, but fail to make full use of the similarity relationship between nodes and their neighbors. Additionally, these methods lack effective utilization of a priori information, making it challenging to extract information about community structure and nonlinear data information in overlapping communities. To address these issues, a method for detecting overlapping communities based on a graph attention autoencoder and self-training clustering (GASTC) is proposed. Firstly, GASTC utilizes the graph attention autoencoder for overlapping community detection. The fuzzy modularity maximization method is embedded into the graph attention autoencoder to perform soft allocation of network nodes. Simultaneously, targeted learning is conducted based on the weights assigned to nodes and their neighboring nodes to capture the interactions between overlapping nodes and different communities. GASTC also designed a structural similarity function suitable for detecting overlapping communities. The community structure within overlapping communities is extracted through a semi-supervised learning approach that not only utilizes label information to enhance the prior, but also introduces connection probabilities between nodes. This enables the calculation of the structural similarity between the known network structure and unlabeled nodes. Finally, subspace clustering is used for self-training, where the cluster labels is used to supervise the learning of potential node features and self-expression coefficient matrices. The obtained self-expression coefficient matrix is used to guide the division of clusters, to capture the non-linear data information in overlapping communities. Experimental results on six datasets demonstrate that GASTC can achieve higher accuracy in overlapping community detection tasks, especially in networks with more complex structures.