View-Consistent Graph Neural Networks

Graphs are common in real life,such as communication networks,social networks,and reference relationships.Because they can naturally reflect the connections between data samples,they have important practical significance and application value for the study of graph data.In recent years,graph neural networks have made a breakthrough in solving large and complex graph processing problems,but there are also issues such as dependence on label information and complex network structures.When there is little or even no label information,the model performance is usually poor.However,in real life,the acquisition of label information is often time-consuming,laborious,and expensive,so how to design a simple and efficient graph neural network model under the condition of very low label rate or even no labels has become a new research problem in the graph learning domain.This paper aims to solve the problem of graph semi-supervised learning with a very low label rate and graph unsupervised learning with no labels from the perspective of multiple views.(1)This paper proposes a general framework of “View-Consistent Graph Neural Networks” based on multi-view networks.The framework can be divided into three parts: a two-branch network,consistency supervision,and task-related supervision.The two-branch network can obtain complementary multi-view representations.Consistency supervision makes multi-view representations learn from and supervise each other.Task-related supervision enables multi-view representations to learn to obtain information relevant to downstream tasks.(2)A semi-supervised View-Consistent Graph Neural Network is designed to solve the semi-supervised classification task of graph nodes.The model learns better sample representations by aligning the semantic consistency of two branch network views,and a novel pseudo-label training strategy is designed to solve the problem of lack of model supervision information caused by insufficient labels.(3)An unsupervised View-Consistent Graph Neural Network is designed to solve the unsupervised representation learning task of graph nodes.The model achieves efficient node representation learning through view consistency supervision and an improved graph contrastive loss function,and extracts single view information and aggregates multi-view information using graph diffusion.In conclusion,the two models designed in this paper can integrate the complementary information of multiple views and better eliminate the graph neural network’s excessive dependence on label information,thus realizing efficient graph node classification and representation learning under the condition of an extremely low label rate or no label.In this paper,relevant experiments were conducted on nine multi-view datasets,including graph node classification experiments under semi-supervised conditions and graph node representation learning experiments under unsupervised conditions.The results show that,under the same experimental conditions,the proposed models have better performance than the baseline models.Furthermore,numerous ablation comparison experiments and parameter sensitivity tests were designed to verify the rationality of the model.