Incomplete Graph-oriented Graph Representation Learning Method

With the development of economy and technology,hundreds of millions of data can be generated every day,and it is one of the hot spots to discover meaningful laws and information from massive data.By using continuous and low-dimensional vectors to represent the network,we can analyze and mine the network more effectively.Capturing the topological structure information or attribute information of the network can obtain the vectored representation of the network.However,the problem of data missing always exists,and it causes deviations in the vector representation of the network,which affects the mining effect of downstream tasks and becomes an urgent problem to be solved.There are various networks in the real world.The lack of data results in the loss of nodes,edges or node attributes in the network.Most of the current research aims the situation of missing edges in the network,rarely involving the loss of nodes and node attributes.In this paper,two common cases of node missing and node attributes missing in the network are studied in depth.The DINE framework and GLAM framework are proposed to solve these two kinds of problems,so as to effectively represent and learn the incomplete network.(1)DINE,the graph representation learning framework for missing network structure: The framework is divided into two parts: network structure recovery and network representation learning.By generating graph model and maximum the expectation model,the missing network structure is estimated to complete the incomplete network.For the recovered network,we propose a network representation learning algorithm based on auto-encoder named MVCDNER,which uses network structure information and node attribute information to perform representation learning on the network to obtain vector representations of nodes.Finally,we perform node classification and link prediction tasks on three datasets of Citeseer,DBLP,and BlogCatalog.The experimental results show that DINE framework outperforms comparison algorithm.As the proportion of missing nodes increases,the results did not fluctuate much.It proves the effectiveness and superiority of DINE.(2)GLAM,the graph representation learning framework for attribute missing: The framework is divided into two parts: node attribute restoration and network representation learning.In the attribute recovery part,we combine the auto-encoder and generative adversarial network to find the mapping relationship between the network structure information and the node attributes,so as to generate the corresponding node attributes from the network structure information and complete the missing attributes of the nodes.In the network representation learning part,we use graph convolutional neural network to learn the representation vector of the network according to the network structure information and node attribute information.Finally,we perform node classification and link prediction tasks on the three datasets of Cora,Pubmed,and BlogCatalog.The experimental results show that the GLAM framework is superior to the comparison algorithms in each evaluation index.As the proportion of missing nodes increases the results keep basically unchanged,which proves the effectiveness of GLAM framework.