Research Of Link Prediction Method Based On Community Structure

Link prediction is one of the research hot spots of social network analysis,which purpose is to predict the missing links or the possible links in the future based on the observed network.Link prediction has extensive application and theoretical value in bio-network analysis,social network recommendation,transportation network planning and so on.At present,many link prediction indices in complex networks using network's topology structural features,but this indices are too simple.On the one hand,it does not consider the community structure of the network,and on the other hand,it does not consider the combination of different network structural features.Based on this,this paper studies the network link prediction problems from the following three aspects.First,the neighbor set information-theoretic model for link prediction does not incorporate the degree of nodes and the community structure of real networks.Therefore,we add the function of node degree to improve link prediction performance.Furthermore,we encode the community information to the prior probability of a node pair being connected,and based on the assumption that the prior probability of two nodes within the same community being connected is greater than that from distinct communities,the neighbor set information indices based on the degree of nodes and the community structure are proposed.The experimental results on a series of real networks show that our methods outperform other classical link prediction indices.Second,the na?ve Bayes model considers the common neighbor feature of the network,but the links across two neighbor sets are also an important feature of the network and it can be used to improve the performance of link prediction.In this paper,the na?ve Bayes model is used to combine two network structural features,namely,common neighbors and links across two neighbor sets.Then we propose the neighbor set na?ve Bayes model.In addition,this paper takes the community structure of the neighbor nodes into account and proposes the neighbor set na?ve Bayes model based indices encoding the community information.Compared with the na?ve Bayes model,the prediction performance of our method is greatly improved.Third,with the development of society,the scale of network have been increasing,and researchers have proposed a lot of link prediction methods.Given a network,how to choose a appropriate link prediction method accurately and quickly is a problem needed to be solved.By taking accuracy of the link prediction algorithms as the response variable,network structural features as predictor variables,we make variable selection,thus the main factors that affect the accuracy of link prediction methods are obtained.Furthermore,this paper carry out the research on the classification of the indices based on tree method,and recommend appropriate link prediction indices for networks according to network topology structure.