Link Prediction Based On Fusion Of Complementary Indexes

With the continuous development of the network age,the research of complex networks has been pushed to a new stage.Link prediction is a very important research topic in the research of complex network science.Link prediction aims to use the existing topology information of the complex network to predict the missing or potential edges in the complex network.Link prediction is closely related to people's real life,and it can be applied to various recommended products to provide people with information services.At the same time,link prediction also has certain theoretical significance.For example,it can provide theoretical support for the research of complex network evolution mechanisms.Over the years,a large number of link prediction algorithms or indexes have been proposed.In general,different link prediction indexes have different advantages and disadvantages.Moreover,different link prediction indexes usually emphasize different aspects of information available in complex networks.Therefore,it is already a challenging and important research topic to fuse different link prediction indexes in different combination ways and play a better role in prediction.In this paper,the fusion of different link prediction indexes is studied to find the complementarity of different prediction indexes and realize efficient prediction.Firstly,the structural perturbation method(SPM)and the random walk with restart(RWR)are fused to obtain a new index with better predictive performance.Then,the linear optimization(LO)method and the resource allocation(RA)index are fused to obtain a new index.Finally,a new link prediction algorithm of directed networks based on the fusion of SPM and LO is proposed,and a new prediction index of directed networks with excellent performance is obtained.The main research contents of this paper are as follows:(1)Link prediction based on structural perturbation method and random walk with restart.Through the research of the SPM,it can be found that the algorithm does not directly use the node relationship information in the link prediction process.However,the SPM considers the overall topology information of the complex network.On the contrary,the RWR is a link prediction index that makes good use of the node relationship information in a complex network.In order to introduce the information of the node relationship into the link prediction process,the SPM and RWR are fused in this paper,and a new index is proposed.Through experimental research,compared with some commonly used classic indexes,it is found that the new index has better predictive performance on most networks.(2)Link prediction based on the fusion of linear optimization and RA index.Through the study of LO,it can be found that LO only uses the information of the odd-numbered paths in the complex network,and lacks the information of the common neighbor nodes in the complex network.Therefore,LO and RA index are fused in this paper,and the information of common neighbor nodes is introduced into the process of link prediction,so as to improve the prediction accuracy of link prediction.The experimental results show that,compared with some classical indexes,the new index obtained by the fusion of LO and RA index makes more comprehensive use of the available information in the complex network,and achieves the higher prediction accuracy.(3)Link prediction based on structural perturbation and linear optimization in directed networks.Many existing link prediction indexes are only applicable to undirected complex networks.However,most real networks are directed networks.For this reason,the fusion of two existing indexes of directed networks is considered in this paper.SPM and LO are fused in the directed networks,and a new index is obtained.Since the perturbation matrix obtained by applying SPM contains the missing edge information in the original complex network,the reconstructed network obtained according to the perturbation matrix contains more network information than the original network.In order to use more network information,LO is further calculated based on the perturbation matrix obtained by SPM.Obviously,the new index obtained after the fusion utilizes more network information,and also fuses the advantages of SPM and LO.The experimental results show that the new index achieves better prediction results than other classical indexes.