Research On Link Prediction By Combining Semantic Structure And Time Characteristics In Heterogeneous Information Network

Link prediction is a research hotspot for data mining and complex networks analyzing.Through the technique of link prediction,the links being formed in the future can be predicted and the current false links can be identified in the networks,which can help us to understand the evolutionary procedure of complex networks.Link prediction has been widely used in many fields,such as friendship relation prediction,e-mail exchanging relation prediction,call relation prediction and cooperation relation prediction.With the development of the Internet,the scale of data is growing dramatically.For current researchers in the field of data mining and complex networks analyzing,it is a research hotspot to apply the technique of link prediction to the environment of large-scale data.The original methods of link prediction were based on homogeneous information network,which required the nodes with the same type and the edges with a single relationship type.So they had some limitations.Though the methods of link prediction based on heterogeneous information network can apply to the networks with different node types and link types,current work only considered limited factors such as topology or semantics.Also most of them focused on the static graph and ignored the time characteristics.Therefore,link prediction for heterogeneous information network is researched in this thesis.A method of semantic structure and time characteristics combined link prediction is presented,which considers a variety of features including topological structure,semantic information and time characteristics to improve the precision of link prediction.The main work and contribution of this thesis are as follows:(1)The related work about link prediction is summarized.And their advantages and disadvantages are analyzed.(2)As for the existing problems,two link prediction models are presented..At first,a semantic structure graph based link prediction model(SLP)is presented,which divides the link prediction into two stages.At the first stage,the semantic structure graph is constructed,which can better reflect the topology structure of the graph,semantic relativity and association strength among nodes.At the second stage,the random walk probabilities are computed and the parameters are optimized.· What's more,based on SLP,a semantic structure and time characteristics combined link prediction model(STLP)is proposed.Different from traditional models,STLP considers a variety of features including topological structure,semantics information and time characteristics.It can improve the precision of link prediction greatly.(3)Based on SLP model and STLP model,two algorithms of link prediction are proposed:semantic structure graph based link prediction algorithm and semantic structure and time characteristics combined link prediction algorithm.Also the strategies of random walk probability computing,parameter optimizing,time similarity measuring and topic attention-degree prediction are discussed.In addition,the algorithms are improved from two aspects:feature choosing in semantic structure graph and jump probability computing,which can further improve the precision of link prediction.(4)The experimental results verify the feasibility and the effectiveness of keytechniques proposed in this thesis.Our SLP method is superior to other link prediction methods on the precision in the whole.Besides,compared with SLP method,STLP and its improved method can improve the precision of link prediction significantly.