基于对比学习的全局增强动态异质图神经网络

Globally Enhanced Heterogeneous Temporal Graph Neural Networks Based on Contrastive Learning

图神经网络由于其对图结构数据的强大表征能力近年来受到广泛关注.现有图神经网络方法主要建模静态同质图数据,然而现实世界复杂系统往往包含多类型动态演化的实体及关系,此类复杂系统更适合建模为动态异质图.目前,动态异质图表示学习方法主要集中于半监督学习范式,其存在监督信息昂贵和泛化性较差等问题.针对以上问题,提出了一种基于对比学习的全局增强动态异质图神经网络.具体地,所提网络首先通过异质层次化注意力机制根据历史信息来生成未来的邻近性保持的节点表示,然后通过对比学习最大化局部节点表示和全局图表示的互信息来丰富节点表示中的全局语义信息.实验结果表明,提出的自监督动态异质图表示学习方法在多个真实世界数据集的链路预测任务上的AUC指标平均提升了3.95%.

Graph neural networks(GNNs)have attracted extensive attention in recent years due to the powerful representation capabilities for graph-structured data.Existing GNNs mainly focus on static homogeneous graph.However,complex systems in the real world often contain multiple types of dynamically evolving entities and relationships,which are more suitable for modeling as heterogeneous temporal graphs(HTGs).Currently,HTG representation learning methods mainly focus on the semi-supervised learning paradigm,which suffers from the problems of expensive supervisory information and poor generalization.Aiming at the above problems,we propose a globally enhanced GNN for HTG based on contrastive learning.Specifically,we use a heterogeneous hierarchical attention mechanism to generate proximity-preserving node representations based on historical information.Furthermore,contrastive learning is used to maximize the mutual information between temporal local and global graph representations,enriching the global semantic information of node representations.The experimental results show that the self-supervised HTG representation learning method proposed in this paper improves the AUC on the link prediction task of multiple real-world datasets by an average of 3.95%.