Meta Path-Based Link Prediction Research For Heterogeneous Information Networks

With the rapid popularization of the Internet,the growing information of network is full of our lives.The traditional single network structure can no longer express the abundant information in the network,so the research of heterogeneous information network is more and more concerned by researchers.The qualitative and quantitative research for heterogeneous information network has become a significant issue in complex network research.Link prediction is one of the primal problems in social network mining.Due to the complexity of the network and the diversity of data,link prediction problems of different types of data also becomes more complicated according to the network structure and existing information in heterogeneous network.Faced with huge information in heterogeneous networks,the existing link prediction technology is not enough to extract and filter data.In addition,link prediction technology in heterogeneous information network mainly considers the structure of the network,which does not fully consider the interrelationships between nodes in the network or too simple to consider.Aiming at the problems of link prediction in heterogeneous information networks,this paper mainly includes the following research contents:(1)Retain the semantic information in the original heterogeneous network and reconstruct the heterogeneous information network,extract the two types of data objects to be predicted in the network and the links between them.Two types of data objects are extracted as node objects in the network,and the number of links between objects is extracted as the link between two types of objects in the network.It simplifies the network,and do not lose the information in the network at the same time.(2)By one kind of objects is considered as the features of the other kind of objects,and the number of links between objects as the eigenvalues,the two types of objects are respectively expressed by vectors and the correlation between nodes in the heterogeneous network is calculated.Among them,the same type of node pairs: cosine similarity between the objects is calculated,different types of node pairs: the ratio between the number of links between two types of objects and the number of all links of an object is calculated,obtain a node correlation matrix model.By constructing the node relation matrix,the semantic information in the heterogeneous information network is fully tapped.(3)In the process of selecting the meta-path,we make full use of the semantic information in the node correlation matrix and propose two rules that make the meta-path include the sub-paths "A-A-B" and "A-B-B",fully tap the network meta-path.By extracting all effective paths in the network,the accuracy of meta-path link prediction in heterogeneous information networks is improved.(4)This paper proposes a meta path-based link prediction research for heterogeneous information networks,called BRLinks.Taking the node correlation matrix as the network structure model,firstly,extract all the instance paths between the two nodes to be predicted as the set of the meta-paths;then calculate the probabilities of the connections between the nodes of each type of meta-paths;and then use the supervised learning method train the weights of each type of meta-path;and finally,integrate the weighted sum of different metapaths between two nodes to link the two nodes.And the validity of the algorithm is verified on the DBLP dataset.Experimental results show that the F-Measure of BRLinks algorithm is higher than link prediction based on common neighbor nodes and restarting random walk.It is of great significance to study the prediction of heterogeneous information network links.