基于自监督信息增强的图表示学习

Graph Representation Learning Enhanced by Self-supervised Information

关键词:  自监督信息  图表示学习  图变分自编码器  图卷积网络  对比损失

Key words:  self-supervised information  graph representation learning  graph variational auto encoders  graph convolutional networks  contrastive loss

摘要:

针对图表示学习模型依赖具体任务进行特征保留以及节点表示的泛化性有限等问题，本文提出一种基于自监督信息增强的图表示学习模型(Self-Variational Graph Auto Encoder，Self-VGAE)。Self-VGAE首先使用图卷积编码器和节点表示内积解码器构建变分图自编码器(Variational Graph Auto Encoder,VGAE)，并对原始图进行特征提取和编码；然后，使用拓扑结构和节点属性生成自监督信息，在模型训练过程中约束节点表示的生成。在多个图分析任务中，Self-VGAE的实验表现均优于当前较为先进的基线模型，表明引入自监督信息能够增强对节点特征相似性和差异性的保留能力以及对拓扑结构的保持、推断能力，并且Self-VGAE具有较强的泛化能力。

Abstract:

Graph representation learning models rely on specific task to preserve features,and the generalization of node representations are limited.Aiming at the above problems,a graph representation learning model Self-Variational Graph Auto Encoder (Self-VGAE) enhanced by self-supervised information is proposed in this article.Firstly,graph convolutional encoder and node representation inner product decoder are used to construct a VGAE. The feature extraction and coding of the original graph are performed.Then,the topology and node attributes are used to generate self-supervised information,and the generation of node representation is constrained during model training.In multiple graph analysis tasks,the experimental performance of Self-VGAE is better than the current more advanced baseline model,which shows that the introduction of self-supervised information can enhance the ability to retain the similarity and difference of node features and the ability to maintain and infer the topology.Furthermore,Self-VGAE has a stronger generalization ability.