基于双重视图耦合的自监督图表示学习模型  

Self-supervised graph representation learning model withdual view coupling

作　　者：陈琪 郭涛[1] 邹俊颖 CHEN Qi;GUO Tao;ZOU Jun-ying(School of Computer Science,Sichuan Normal University,Chengdu 610101,China)

机构地区：[1]四川师范大学计算机科学学院,四川成都610101

出　　处：《计算机工程与设计》2023年第12期3738-3744,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(11905153)。

摘　　要：针对现有的图表示学习在自监督对比学习方法中存在视图差异较大,且依赖于负样本防止模型坍塌,导致节点表示能力弱及空间复杂度加大的问题,提出一种基于双重视图耦合的自监督图表示学习模型(self-supervised graph representation learning model with dual view coupling, DVCGRL),用于学习图数据表示。采用特征空间增广和结构空间扩充相结合生成双重视图,将双重视图作为正样本对输入孪生神经网络;利用图编码器提取图数据特征,通过多层感知器获得映射后的特征向量;采用耦合网络拉近双重视图的特征向量距离,提升节点表示能力,防止模型坍塌。在公开数据集上进行的节点分类实验结果表明,与当前主流图表示学习模型相比,该模型降低了空间复杂度,节点分类精度得到明显提高。To address the problems that existing graph representation learning has large view differences in the self-supervised comparative learning method and relies on negative samples to prevent model collapse,resulting in weak node representation and increased spatial complexity,a self-supervised graph representation learning model based on dual view coupling(self-supervised graph representation learning model with dual view coupling,DVCGRL)was proposed for learning graph data representation.A combination of feature space augmentation and structure space expansion was used to generate dual views,which were fed into the twin neural network as positive sample pairs.The graph encoder was used to extract graph data features,and the mapped feature vectors were obtained by a multilayer perceptron.The coupled network was used to close the feature vector distance of the dual views to enhance the node representation and prevent model collapse.Node classification experiments on publicly available datasets show that the model reduces the spatial complexity and the node classification accuracy is significantly improved compared with the current mainstream graph representation learning models.

关 键 词：双重视图 孪生神经网络 图表示学习 图卷积网络 图数据增广 节点分类 自监督对比学习 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]