Link Analysis And Prediction Of Heterogeneous Information Network

In the real world,there are many complex networks,such as social networks,information networks,and biological networks.Nodes or entities in these complex networks are used to represent person,computer,biological elements,and so on.And links between nodes are used to represent a relationship or interaction between nodes.Analyzing the behavior or relationship of entities in the network is an important research content in complex network research.However,due to limitations of the time,the methods of data collection and the observer,the observation of the network with missing links often occurs.Link prediction is to find missing links and predict possible links in the future through network topology information,node attributes,observed links and other information.Link prediction is widely used in many different fields,for example,to recommend possible friends to users on social networking sites;to predict which researchers are likely to collaborate in the future in scientific networks;to looking for data or records that have been deleted from the network,and so on.Therefore,this paper focuses on the research of link prediction technology,the main research contents are as follows:This paper first studies the link prediction technology for isomorphic topological complex networks.Network embedding(NE)can effectively reduce the data dimension and improve the operational efficiency of complex network link prediction.The existing network embedding researches mainly use the structure information of the network as the embedding features to encode.Although some of these algorithms add text content,these methods without embedding the connection between different nodes,can't express the different semantic relationship when connected with different neighbor nodes.Therefore,this paper proposes a collaborative embedding model(NSTI)with network structure and text information,which not only learns the node embedding based on network structure information,but also learns the interaction information between nodes and different neighbor nodes.Instead of learning the words of a single node,the interaction information between the embedded words of the pair of nodes is calculated,so as to express the different semantic relations when the node is connected with different neighbor nodes.Experiments on link prediction on real data sets show that the proposed model and algorithm improve by 2%-15% in AUC values compared with the existing algorithms such as Deep Walk,LINE,Node2 vec,TADW and CENE,which proves that the model can effectively improve the accuracy of link prediction.In the real world,most systems are heterogeneous information networks,and there are often many types of nodes and many types of relationships between nodes.We then studied the link prediction technology of heterogeneous information networks.The relationship between nodes in heterogeneous information network is more complex than that of isomorphic information network,and it is necessary to deal with the complex relationship between nodes and nodes in heterogeneous information network by meta-path.However,the embedding methods based on meta-paths learning alone does not fully represent the mutual information between nodes in heterogeneous information networks.Therefore,we calculate the similarity of attribute characteristics between node pairs in the network to improve the accuracy of the link prediction.To deal with different semantic information,we propose fusion functions of different meta-paths to improve the accuracy of the embedded vectors of nodes.Based on the above research,a network embedding model(NSMF)for the heterogeneous information network by node similarity and meta-path is proposed.Through the analysis of experimental results on the real two heterogeneous information network data sets,the proposed NSMF model algorithm improved by about 30%-40% in AUC values compared with the algorithms Deep Walk,Node2 vec and LINE for isomorphic information networks.And for several algorithms Hin2 vec,Metapath2vec and PTE for heterogeneous information networks,the AUC value of the NSMF model algorithm was increased by about 2%-10%.The experimental results prove the validity of the model in the link prediction experiment.