Adaptive graph contrastive learning for community detection

Recently, graph contrastive learning (GCL) has received considerable interest in graph representation learning for its robustness in capturing complex relationships between nodes in an unsupervised manner, making it suitable for unsupervised graph learning tasks such as community detection. However, most GCL approaches have two limitations when applied to community detection. First, the random augmentation strategy employed by them may destroy a graph’s community structure due to the random added/removed edges or attributes. Second, nodes with similar topology or attributes may be selected as the negative samples of a target node according to their sample selection strategy, leading to the wrong assignment of the target node’s community. In this paper, we propose an adaptive-graph-contrastive-learning-based community detection (AGCLCD) algorithm to address the problems. At its core, AGCLCD introduces an adaptive graph augmentation strategy to preserve a graph’s original community structure in augmentation. Furthermore, we develop a composite contrastive pair selection scheme to choose the nodes sharing similar topology and attributes with a target node as its positive samples to ensure that the representation vectors of nodes in the same community are highly relevant. Comprehensive experiments on real-world and synthetic networks demonstrate that AGCLCD achieves higher accuracy and effectiveness than state-of-the-art algorithms.