Research On Link Prediction Algorithm For Social Networks

With the rapid development of information technology,social applications have become an integral part of people's lives,and formed huge social networks which include a wide variety of information.As an important research area of data mining in social network,link prediction in social network can help researchers to assess the network formation mechanism,it can also be used to predict missing information and dynamic change of the network.It can help users to discover the content that they may be interested in which is useful to deal with the problem of information overload.Therefore,the research of link prediction in social network has great significance for academic research and commercial application which attracts more and more attention of researchers.However,there are many difficulties which are not be solved currently.Firstly,social network is built based on people 's behavior,there are many factors will affect the change of links between nodes in social network because individuals are not completely independent,all of factors should be studied when we predict links.Secondly,social networks contain a lots of nodes and information,if we predict links based only on similarity between nodes,we cannot make full use of the information which will have negative effects on precision of link prediction algorithms.Therefore,with increase requirements of users and expansion of network,how to fully study the information in social networks for link prediction has become an important research content.According to the research status of link prediction for social networks,the research contents of this paper are as follows:(1)Whether the link is established or not is related to centrality of nodes,we studied link prediction algorithm combining centrality of nodes,and proposed a new algorithm named LRC to evaluate the centrality of nodes.The traditional local centrality have not taken influence between nodes account.According to nodes' egocentric network,we proposed relational weight to describe the mutual influence between adjacent nodes.Finally,centrality of nodes is evaluated based on its position and the state of its egocentric network.Experiment show that LRC algorithm can find essential nodes effectively which is more accurate than those classical local centrality algorithms.(2)For research on link prediction for social network,the traditional similarity-based link prediction algorithm does not adequately consider the relationship between nodes and the characteristics of social networks,we studied link prediction algorithms based on improved similarity.We proposed the algorithm named LP-LRC based on LRC which predicts links based on weak ties theory of social networks and LRC of nodes,it improves the effect of common neighbors with low centrality and reduces the effect of common neighbors with high centrality,which is more suitable for social network.We proposed a link prediction algorithm named RWCN which is an improved algorithm based on CN,it differentiates the impact of different common neighbors on the establishment of new links based on relational weight between nodes.The experimental results show that our algorithms have achieved better performance than those traditional algorithms and some improved algorithms.(3)For social networks that contain domain related information,we studied link prediction algorithm based on multi-features.Two methods to extract features based on domain information are proposed to describe the degree of activity of nodes and the similarity of interest between nodes.We built feature vectors to predict links by combining features based on topological structure and domain related information.The experimental results based on DBLP dataset show that domain features we extracted have higher information gain than other features and have a positive impact on the performance of the algorithm.