| The purpose of network representation learning is to map high dimensional sparse network to low dimensional dense real valued vector space.It can adaptively extract features and facilitate the analysis of downstream network tasks.The network representation method for non-attribute graphs can use the vector of network to gain insight into social structure,explore language communication mode and discover complex network structure.The existing related work is mainly to input the network structure into the shallow random walk model or deep graph neural network model to obtain the vector of network nodes,and apply it to the downstream analysis task.At present,there are some difficulties in the research:(1)the current shallow random walk model in network representation learning cannot effectively capture the similar characteristics of node structure and content at the same time;(2)the deep graph convolution neural model cannot be better embedded in the non-attribute graph in an unsupervised way,and its applicability is poor on the graph;(3)the improved deep graph convolution network(HGCN)model mainly captures the nonlinear structure by aggregating the characteristics of neighbor nodes,ignoring the linear structure of the network.We proposes a model(SN2vec),which integrates node structure and content to solve the problem of poor performance of shallow random walk model on the network data with equivalent mix of structure and content;proposes an unsupervised graph embedding model(HGCN)based on hierarchical graph convolution network to solve the problem of poor applicability of deep graph convolution neural model on non-attribute graphs;and proposes a network representation method(HGCN-DL)embedded with linear and nonlinear structure.The main work of the paper includes:Firstly,a network representation method integrating node structure and content is proposed.The shallow random walk model in network representation learning mainly includes two kinds of methods: node structure similarity and content similarity.In order to overcome the shortcoming that the existing methods cannot capture node structure and content effectively at the same time,we redefines node similarity and proposes a joint representation method of node structure and content similarity(SN2vec).The experimental results of node classification and dimensionality reduction show that the proposed method can capture the similar information of node structure and content more effectively.Secondly,an enhanced unsupervised graph embedding model(HGCN)is proposed.Graph convolution network(GCN)is a semi-supervised learning model,which integrates label information into the learning process of network structure.However,the existing network data attribute is not accurate and the attribute is not easy to obtain,which leads to the poor applicability of graph convolution network on the non-attribute graph.Therefore,the paper improves the method of obtaining the attribute and label of network nodes,and improves the applicability of graph convolution network on non-attribute graph.Experimental results on different kinds of data sets show that the proposed method improves the applicability of graph convolution network on non-attribute graph.Thirdly,a network representation method embedded with linear and nonlinear structures is proposed.In order to overcome the problem that unsupervised hierarchical convolution network(HGCN)model captures linear structure,the local and global linear structures of the network are defined.And we propose a network representation model HGCN-DL,which embeds both linear structure and nonlinear structure.The experimental results of node classification and dimensionality reduction show that HGCN-DL model can capture the linear and nonlinear structure of network more comprehensively. |
