Link Prediction Algorithms Based On Improved Local Na?ve Bayes Model

With the rapid development of various information technologies represented by the Internet,mankind has entered the network age.Today,people live in a world full of various complex networks,and the network science that takes complex networks as research objects has also developed rapidly.As one of the hot research directions in the network science,link prediction aims to predict the possibility of the existence of a link between a pair of disconnected nodes in a network based on information such as the topology and node attributes of the network.It includes predictions of missing links and future links.The research of link prediction not only has extensive practical applications,but also is significent in theory.At present,researchers have proposed a large number of link prediction methods from different perspectives.Among them,the similarity method based on network topology has attracted much attention.The Local Na?ve Bayes(LNB)model is an efficient one which has higher prediction precision and relatively lower computational complexity.This paper proposes an improved LNB model.The LNB model suggests that different common neighbors have different effects.It discriminates the contribution of different common neighbors by a role function and hence improves the accuracy of prediction.However,the LNB model ignores the influence of other structural information.Inspired by the CAR method,this paper considers the influence of local community structure on the basis of the original LNB model.A new role function is introduced in the improved LNB model,which adds the contribution of local community links to the prediction results.Then the impact of the degree distribution of candidate node pairs on the prediction results.The connected links in the network are grouped according to the degree distribution.The more connected links is in a certain group,the more the network tends to produce such links.According to this partition,we can determine the probability of the existence of a link between two unconnected nodes based on their degrees.In order to verify the prediction accuracy of the proposed model,we performed experiments on 12 real networks.The experimental results show that our model achieves the best prediction performance in most cases,and has a significant improvement over the prediction accuracy of the original LNB model.Based on the work of unweighted networks,we extend the proposed model to weighted networks.We first make further improvements to the role function in the LNB model by considering the influence of the link weights between candidate nodes and common neighbors.Then the clustering coefficient in the original LNB model are replaced by the weighted clustering coefficient that is more suitable for weighted networks.Finally,we replace the degree of nodes considered in the unweighted network with the strength of nodes.Because classifying the links based on strength may make the groups of links very large,the number of links in each group is very small.Therefore,we divide the strength of nodes into equal-width intervals,and then classify the candidate node pairs according to the interval which the strength of the candidate node pairs belong to.The experimental results show that the model we proposed for weighted networks achieves excellent prediction performance,and also verify the improvement of prediction performance in weighted networks.