Research Of Link Prediction In Social Network Based On Probability Model

Link prediction is one research direction in the field of social network analysis, which purpose is to predict the probability of the link between the non-connected node pairs. Link prediction has a wide range of applications in fields of sociology, bioinformatics, e-commerce and so on.At present, the most of works has been done, which define similarity indices based on network’s topology(such as: CN, AA, LHN-II and so on), but these methods are too simple to depict deep-seated structural information of network. Therefore, on the basis of the simple unweight network, this paper discusses the problem from the following two main aspects:First, aiming at the weak link in the Na?ve Bayes model independence assumption, we introduce the Hidden Na?ve Bayesian model. Meanwhile, we combine the characteristics of Hidden Na?ve Bayesian model and the connectivity of local network, which makes up by common neighbors of link pair of nodes. In the process, we intensify the relationship between them by mutual information, and get link prediction algorithm based on Hidden Na?ve Bayesian model. The experimental results on four real-world social network data sets Ca-GrQr, Facebook, Eron and Advogato show that Hidden Na?ve Bayesian Link prediction has higher AUC and Precision than Na?ve Bayesian Link prediction model.Second, we analyses similarity indices that introduced above roundly form the qualitative point of view, and consider them as the structure architectural of the link in the network. Inspired by these features, a new Cohesion index is constructed, and the validity of this feature is illustrated by a comparative experiment. On this basis, from the perspective of machine learning, these features are selected, and the performance evaluation system of feature selection based on classifier performance evaluation index is established. We solve the imbalance problem in the process of feature selection and effect evaluation by using the method, which does KMEANS cluster first and sample according to the result of cluster. On Twitter, Facebook Jazz and Email four data sets to do the experiments. The experimental results show that the proposed method and model are effective. In the evaluation system five indicators ACC, REC, PPV, 1F-score and AUC have better performance overall.In addition, based on the above research, this paper extends the Na?ve Bayesian link prediction model to the weighted network. This expansion is mainly set up in social network link structure feature selection model frame. And we focus on the performance of different classifiers when included weighted characteristics. The experiment shows that this expend has some feasibility.