Link Prediction In Social Networks Based On Improved Similarity

Link prediction is a hot issue in the field of social networks analysis. It aim s at estimating the likelihood of the link formulation between nodes based on network structural information. Currently, most link prediction algorithms depend on the network characteristics. Through mining the network structure features leads to calculate the similarity between nodes, this type of algorithms judge whether the link is being. The representable algorithms base on the information of common neighbors, such as CN, AA, RA and so on. But the number of co-neighbor node s and degree of co-neighbor node s in t his type of algorithm cannot fully reflect the complex structure of social networks.Link prediction algorithm based on Bayesian has good predictive results. But the algorithm has two drawbacks:(1) the assumption of independence, and the assumption is not true in practice;(2) it bases on the u nweighted network, which ignores the weight of social networks.According to the above analysis, we study the problem of link prediction algorithm. The detail research process and major works are as follows:(1) In order to dig more information of structural properties in the network, we propose a new algorithm called CN-based Enhanced algorithm(E_CN) by combining with the concept of community division in the third chapter of paper. N odes are divided into different communities by this new algorithm and each node has its role function. Simultaneously, we apply the new algorit hm to Jaccard. Finally, we do the experiments on the m anual data sets and real data sets, and we obtain that the new algorithm is better than the original classic algorithms.(2) To overcome the first drawback of LNB model, we introduces a Tree Argument Bayesian model to relax the assumption of independence in the fourth chapter. Based on TAN model and combined with entropy and CN algorithm, we get a new algorithm TAN_CN. We apply the TAN model to AA, RA, and get TAN_AA, TAN_RA. Finally, we do some experiments in the real data sets, and we obtain that the algorithms of TAN model are better than the algorithms of the classic model and the algorithms of the LNB model.(3) To overcome the second drawback of LNB model, our study expands to the weighted network. We propose CN, AA, RA of the WLNB model, which adds in Clustering Coefficient. The algorithms based on WLNB model reflect the effect of weighted structure of the co-neighbors in the similarity, and their accuracy improves. Finally, the algorithms are applied to unweighted networks, and weighted algorithms also achieve a free c onversion between unweighted network and weighted network.