Hyperbolic distance guided multi-positive graph contrastive learning

Graph contrastive learning (GCL) aims to obtain embeddings for graph data in an unsupervised manner. Most existing GCL models are designed for homogeneous graphs, whereas the two critical aspects of heterogeneous GCL models are data augmentation and contrastive sample selection. In this paper, we propose a contrastive learning model for heterogeneous graphs. We take advantage of the property that hyperbolic space is suitable for representing highly hierarchical graph data. We use the normalized hyperbolic distance as an indicator to adaptively select edges to remove in data augmentation to obtain different graph views. In addition, we propose a positive example expansion strategy that integrates topological information and attributes information to obtain more highly relevant positive samples and provide more contrasting information. We conduct node classification experiments on several public heterogeneous graph datasets to evaluate the performance of our model. The results show that our model outperforms all unsupervised comparison methods and even outperforms the best-supervised model on several datasets to reach state-of-the-art.