Research On Network Representation Learning Based On Structural Proximity

With the rise of social service systems such as Facebook,Twitter,We Chat,Weibo and Douban,not only unlimited communication between people has been realized,but also information exchange has been realized.With the continuous development of information technology,the scale of users in all kinds of systems continues to expand,forming one large-scale information network after another,at the same time,the accumulation of massive and rich data.How to excavate the potential valuable information through various network analysis tasks in these massive data is an important problem to be solved.Network representation learning,also known as network embedding or graph embedding,is a fundamental and core problem in network analysis task.It has been widely paid attention by industry and academia,and has become one of the research hotspots.In this thesis,representation learning is studied in social network and signed network based on the method of optimizing objective function.The specific research contents are as follows.Firstly,a Structural Proximity based social Network representation learning model,SPNE(Structural Proximity Based Network Embedding),was proposed to solve the disadvantages of the LINE(Large-scale Information Network Embedding)model.In order to solve the problem of the same empirical probability of vertices in LINE first-order proximity,the preprocessed Katz index is introduced as the empirical probability.In order to solve the problem that the first-order and second-order proximity are independent of each other in LINE model,a new joint probability of local and global structural information of the fused network is constructed.By minimizing joint probability and empirical probability,more accurate vector representation of social network vertices is realized.Secondly,the SPNE model was extended from social Network to signed Network,and a signed Network representation learning model based on Structural Proximity--SPSNE(Structural Proximity Based Signed Network Embedding)model was proposed.In order to distinguish the influence of positive and negative signs between vertices on the structure neighborhood,the definition of first-order proximity and second-order proximity of the signed network is proposed based on the structural balance theory.In order to make the extended model fit for the signed network,the joint probability,empirical probability and loss objective function are constructed respectively for positive and negative edges,and the two loss objective functions are fused as the total loss objective function.Finally,high quality signed network representation learning is achieved by minimizing the optimization loss objective function.Finally,in the social network and signed network,three classical comparison algorithms and three data sets are selected to carry out comparative experiments.The accuracy and accuracy of link prediction based on network representation learning of SPNE model or SPSNE model are verified.