Leveraging Explicit Local Semantics For Graph Representation Learning

Graph representation learning aims to learn the representation of objects based on their own features and correlations with others.It is commonly used as end-toend object classification,or downstream machine learning tasks,such as prediction,quantification,visualization,and reasoning.Deep graph representation learning can be divided into two types: semi-supervised graph learning use parameterized matrices to propagate labeled information,which faces the over-smoothing problem.Unsupervised graph learning is based on the assumption of label smoothness,the strength of node correlation depends largely on the selection of similarity function.In this paper,we introduce local semantics to improve the above problems.The contributions of this paper are as follows:First,for semi-supervised graph representation learning,we propose a parallel training method combining unsupervised and semi-supervised information,which can effectively measure the unsupervised local semantics,and maintain both unsupervised and semi-supervised information during the training.In addition,a variant of crossentropy function is proposed to deal with the over-smoothing of graph convolution networks.Second,for unsupervised graph representation learning,we propose a preprocessing method based on Gaussian embedding,which divides the original graph into path and subgraph set.Through the special property of path,path embedding is proposed to embed the local semantics of path into the similar Gaussian distributions.This method makes full use of the local and global semantics of nodes and improves the performance of unsupervised Gaussian graph learning.Experiments in several datasets and baseline models show that leveraging explicit local semantics can improve the performance of semi-supervised and unsupervised graph representation learning,which further reflects the advantages of explicit integration of local semantics into graph representation learning model.