Heterogeneous Graph Contrastive Learning with Dual Aggregation Scheme and Adaptive Augmentation

Heterogeneous graphs are ubiquitous in the real world, such as online shopping networks, academic citation networks, etc. Heterogeneous Graph Neural Networks (HGNNs) have been widely used to capture rich semantic information on graph data, showing strong potential for application in real-world scenarios. However, the semantic information is not fully exploited by existing heterogeneous graph models in the following two aspects: (1) Most HGNNs use only meta-path scheme to model semantic information, which ignores local structure information. (2) The influence of cross-scheme contrast on the model performance is not taken into account. To fill above gaps, we propose a novel Contrastive Learning model on Heterogeneous Graphs (CLHG). Firstly, CLHG encodes local structure and semantic information by a dual aggregation scheme (i.e. network schema and meta-path). Secondly, we perform contrast between views within the same scheme and then comprehensively utilize dual aggregation scheme to collaboratively optimize CLHG. Furthermore, we extend adaptive augmentation to heterogeneous graphs to generate high-quality positive and negative samples, which greatly improves the performance of CLHG. Extensive experiments on three real-world datasets demonstrate that our proposed model achieves competitive results with the state-of-the-art methods.