Research On Link Prediction Algorithm Fusing Network Topological Structure Features

As one of the hot research directions of complex networks,link prediction plays an important theoretical and practical value in the biological field,social field,traffic field and other scenarios.Research on efficient link prediction algorithm is a very practical work.The goal of link prediction algorithm is to predict the links that are not observed or may appear in the future through the existing node information or topology structure information(such as node attribute,link weight,community structure,etc.).However,most existing methods only consider local or global features,and use limited network topology information,which affects the performance of the link prediction algorithm.In order to solve the above problems,a symmetric non-negative matrix factorization link prediction algorithm combining the local topological features of the network is proposed by using the symmetric non-negative matrix factorization model.Then,a link prediction algorithm based on community closeness is proposed by using the community information of nodes to calculate the relationship between different communities in the network and combining the local similarity indexes such as common neighbors.The main work of this paper is as follows:(1)Aiming at the problem that the SNMF method can't capture the local information of the network effectively,a symmetric non-negative matrix factorization link prediction algorithm(LNF-SNMF)integrating network local topological features is proposed by using the feature matrix constructed from the network local topological information as the input of SNMF model.Firstly,the SNMF model based on the ? and ? update rules is applied to the link prediction algorithm,and the local topological feature matrix of the network is constructed by using the clustering and diffusion characteristics of the link edges and the common neighbor information of the nodes as the input of the model.Thus,two methods based on SNMF-? and SNMF-? are obtained.In order to evaluate the performance of these two methods,this paper uses Karate and other real networks in six different domains to conduct experiments,and compares them with nine existing link prediction algorithms.Experimental results show that the prediction accuracy of the two proposed methods is better than that of the traditional link prediction algorithms based on non-negative matrix factorization.(2)Aiming at the problem that the existing similarity methods can not effectively mine the community structure information,a link prediction model based on community intimacy(CI) is proposed,which considers the relationship between different communities.In this paper,two community detection algorithms,Louvain and FastQ are used to divide the communities.Then,the community intimacy is calculated according to the community structure of the communities where the nodes are located,and the probability of mutual connection between two unconnected nodes is calculated by combining the algorithms based on common neighbors.Finally,experiments were carried out on 10 real networks such as USAir.The experiments showed that compared with four baselines such as CN index,the AUC index was up to 2%,and the ranking score index was up to 10%,indicating that the performance of link prediction could be improved by using community structure information.