Research On Link Prediction Technology In Graph Data

The world of matter and everything that human being are facing is not only related but also developed forever,and understanding the law of development and change of things has always been a hot topic of scientific research.People usually depict and describe complex systems in the real world by means of graphs.There are different types of objects and different types of links in a complex system,and each object and link contains different meanings.Understanding and mastering the hidden information in graph data is an important part of data mining research.Due to various reasons,people cannot completely obtain the information of various objects and links in a complex system,instead,we need to analyze and judge implicit links,missing links and false links based on known information.Therefore,link prediction can give us a better understanding and analysis of complex systems.Researchers have studied the link prediction in different fields,including computer science,physics,sociology,biology,and so on.Various methods were proposed and mainly based on the structural characteristics of the network,Markov chain and machine learning methods.The core of various link prediction methods is the similarity measure between objects.Researchers have proposed a variety of similarity indexes to measure the similarity of objects from different perspectives.Most of the current studies are based on similarity of nodes,path similarity and other indicators.Therefore,it is very important to measure the similarity by the structure information of the network and the attribute information of the node in the graph data,which directly affects the quality of the link prediction.At present,there are two problems in the study of link prediction:(1)How to improve the accuracy of similarity measure.The current similarity measure only considers the number of common neighbors and ignores all neighbors of individual.Moreover;it usually considers the structural similarity and ignores the similarity of node attributes.This will cause the similarity measure accuracy to be not high enough.(2)Link prediction in heterogeneous networks.The traditional link prediction model is mostly used in homogeneous networks;the prediction accuracy is generally not high enough while used in heterogeneous networks,This paper focuses on the above two issues and studies in three parts.Combining global similarity and local similarity,it proposes a method for link predicting in graph data.The main contents and results of this paper are as follows:(1)For the similarity measurement problem,we have studied the asymmetry information of similarity between objects and proposed a similarity measure based on similarity elements(SMSE)in this paper.The traditional similarity measure considers that if the two nodes are similar,the degree of similarity is the same.But it has only considered the number of common neighbors while ignored the neighbors of individual.The SMSE method considers the asymmetric information of individual neighbor's number.Compared with the traditional similarity measure,SMSE method can better distinguish the similar between two nodes.The experimental results show that the SMSE measure can improve the accuracy of the similarity measure and help to improve the accuracy of link prediction based on similarity.(2)The traditional similarity measure emphasizes the information o network structure while elects the information of attribute.We treat the similar information of the network structure as globally similar information and the similar information of the node attributes as local similar information.The Object-oriented Similarity Algorithm(OSA)which combines global similarity and local similarity is proposed.It takes into account structural information and attributes information,avoiding the one-sided consideration of structural similarity.Experimental results show that the OSA method proposed in this paper improves the accuracy of the similarity measure.(3)Link prediction problem in heterogeneous networks.We use node2vec's network representation learning method to collect samples of neighbor nodes in heterogeneous networks and it can preserve the global and local information effectively.Then we calculate the similarity in heterogeneous network by HeteSim measure and build a Link Prediction in Heterogeneous Network(LPHN)model.The experimental results show that the LPHN method proposed in this paper can effectively reduce computational complexity of similarity and provide a new method for link prediction in heterogeneous networks.In summary,the method proposed in this paper improves the accuracy of the similarity measure and the accuracy of link prediction.It can be widely used in various fields.