基于自监督学习的不平衡节点分类算法

Imbalanced Node Classification Algorithm Based on Self-Supervised Learning

在现实世界的节点分类场景中,只有少部分节点带标签且类标签是不平衡的.然而,大部分已有的方法未同时考虑监督信息缺乏与节点类不平衡这两个问题,不能保证节点分类性能的提升.为此,文中提出基于自监督学习的不平衡节点分类算法.首先,通过图数据增强生成原图的不同视图.然后,利用自监督学习最大化不同视图间节点表示的一致性以学习节点表示.该算法通过自监督学习扩充监督信息,增强节点的表达能力.此外,在交叉熵损失和自监督对比损失的基础上,设计语义约束损失,保持图数据增强中语义的一致性.在三个真实图数据集上的实验表明,文中算法在解决不平衡节点分类问题上具有较优的性能.

In real-world node classification scenarios,only a few nodes are labeled and their class labels are imbalanced.In most of the existing methods,the lack of the supervision information and the imbalance of node classes are not taken into account at the same time,and the improvement of node classification performance cannot be guaranteed.Therefore,an imbalanced node classification algorithm based on self-supervised learning is proposed.Firstly,different views of the original graph are generated through graph data augmentation.Then,node representations are learned by maximizing the consistency of node representations across views using self-supervised learning.The supervised information is expanded and the expressive ability of nodes is enhanced by self-supervised learning.In addition,a semantic constraint loss is designed to ensure semantic consistency in graph data augmentation along with cross-entropy loss and self-supervised contrastive loss.Experimental results on three real graph datasets show that the proposed algorithm achieves better performance on solving the imbalanced node classification problem.