Research Of Link Prediction In Social Networks

Nowadays,with the rapid development of the Internet,people have been used to transmit information through the Internet.The increasing number of users and complex relationships constitute a colorful online social network.As an important carrier of people's online social activities,social network has the characteristics of huge amount of data information,diversity of communication and complex network structure.It leads to the continuous evolution of network structure which is difficult to control.It can not wait to do some research about grasping the evolution trend and development rule of social network,excavating the potential association of users and predicting the link relationship between users effectively.This thesis mainly studies the link prediction in social network from user attribute and network structure.The contribution of the thesis can be summarized as follows:1.In the analysis of user attribute,considering the coupling relation between multidimensional attribute factors,a correlations multiplexing for link prediction model in multidimensional networks spaces can be proposed based user behavior and relationship.Firstly,based on hierarchical theory,the network can be mapped into multidimensional networks spaces.The coupling relationship between the spaces can be reduced and the correlation can be analysed in each space separately.Secondly,by considering the characteristic of user multiple interests and taking the advantage of Latent Dirichlet Allocation(LDA)with the problem of interest partition,the traditional text modeling is applied to user behavior modeling.Moreover,the method can be extended with Weighted Average of One-Dependence Estimators(WAODE)which is good at reducing attribute independence.And multidimensional factors can be combined to link prediction.The link prediction model not only can be used to link prediction,but also can be used to find the key factors in link establishment.2.In the analysis of network structure,taking full account of the correlation between the non topological features and the topological features,a three-level hidden Bayesian link prediction model in social network is proposed.Firstly,by applying Gaussian weighted LDA to user behavior modeling,user latent interest distribution can be mined and the influence of internal driving factors can be analysed.Secondly,by introducing user activity to further optimize the Gaussian weighted LDA,the interpretability of the model can be enhanced and the expressive ability of the interest topic can be improved.Moreover,based on latent interest distribution,the implicit factor of Hidden Naive Bayes(HNB)algorithm is redefined.By analyzing the influence of external driving factors,the improved HNB algorithm can be used to link prediction.The link prediction model not only can be used to mine user latent interest distribution,but also the performance of link prediction can be improved.At last,the thesis verifies the proposed models by using some data form Sina Weibo and Twitter.Experimental results indicate that the proposed models can effectively combine user attribute and network structure and predict the link among the users.In addition,the thesis can be beneficial to master the evolution trend and the development law in social network,and it also can provide reference value for the recommendation.