Research On Self-supervised Attributed Network Representation Learning

Attributed networks are widely distributed in the real world,including citation networks,social networks,and transportation networks.Attributed networks contain rich information,including complex relationships between nodes,attributes carried by nodes,and global properties of attributed networks.Due to the complexity and diversity of attributed networks,direct analysis of attributed networks requires not only expert knowledge but also a lot of attempts.Attributed network representation learning aims to map the nodes to low-dimensional representations while preserving the rich semantic information and the proximities.Since the size of the attributed network is usually very large in the real world,manual labeling of these nodes costs a lot.Therefore,how to learn node representations without relying on manual annotation information is a hot research direction in recent years.This thesis focuses on research on self-supervised attributed network representation learning,aiming to automatically extract supervision signals from the data itself,and train the model to obtain high-quality node representations.The nodes in the attributed network have two different kinds of information,namely topology information and attribute information,wherein topology information can be divided into low-order structural information and high-order structural information.Existing methods can capture low-order structural information well,but cannot model the high-order structural information of nodes well.At the same time,the strategy of fusing node topology information and attribute information in existing methods are simple,and cannot capture the correlation between the two very well.The methods proposed in this paper are property prediction-based and contrastbased self-supervised attributed network representation methods.The research contents and main contributions are as follows:First,for the problem that existing methods cannot well model the high-order structural information,a self-supervised attributed network representation learning method based on property prediction was proposed.The proposed method extracts the positive pointwise mutual information matrix as high-order structural information.The proposed method predicts the high-order structural information from adjacency matrix and node attributes,so as to fuse the attribute information of the node and various structural approximations.The proposed method can learn high-quality node representations.In addition,the proposed method uses the principal component analysis method to reduce the dimension of the positive pointwise mutual information matrix,which greatly reduces the space complexity of the model.Second,for the problem that existing methods cannot well integrate node topology information and attribute information,a contrast-based self-supervised attributed network representation learning method was proposed.The proposed method maximizes the mutual information between node topology representation and the attribute representation,thereby modeling the correlation between node topology information and attribute information,and fusing the complementary information of the two.The proposed method does not rely on augmentation operations and graph convolutional network.In addition,the fusion strategy proposed in this method can flexibly select different topology information as model input.The proposed method achieves excellent performance on node classification and link prediction tasks,which proves the effectiveness of the fusion strategy of the proposed method.