基于对比学习增强的图卷积网络

Graph Convolutional Networks Enhanced by Contrastive Learning

多数图卷积模型通过引入更高效的信息传递和转换方式提升图分析任务的实验表现,忽视了原始图信息的保留.针对上述问题,本文提出了一种同时保留节点属性和拓扑结构信息的对比损失函数,增强图卷积模型的表征能力.该方法首先构建基于节点属性和拓扑结构的连接强度,然后使用对比损失保留连接强度中蕴含的属性和拓扑关联信息,最后构建引入交叉熵损失的多层图卷积模型进行端到端训练.在性能提升实验和对比损失比较实验中,引入对比学习的图卷积模型始终优于基线以及其他对比学习方法.实验结果表明,属性和拓扑对比损失能够增强模型对原始图信息的保留,有效提升图卷积模型在节点分类任务中的实验表现.

Most graph convolutional models improve performance by introducing efficient information propagation and aggregation methods,and ignore to preserve original graph features.Aiming at the above problems,this paper proposes a contrastive loss function to preserve attribute and topology information,and enhanced the representation ability of GCN.Firstly,node attributes and topology are used to generate connection strength.Secondly,contrastive loss was used to preserve the attribute and topology information contained in connection strength.Thirdly,the multi-layer graph convolutional models were built,and the cross-entropy loss function was introduced for end-to-end training.In the performance improvement and contrastive loss comparison experiments,the graph convolutional model with contrastive learning consistently outperforms the baselines and other contrastive learning methods.The results show that attribute and topology contrastive loss can retain richer graph information,and significantly improve the performance of the graph convolutional models in node classification.