Hidden link prediction based on node centrality and weak ties

Link prediction has been widely used to extract missing information, identify spurious interactions, evaluate network evolving mechanisms, and so on. In this context, similaritybased algorithms have become the mainstream. However, most of them take into account the contributions of each common neighbor equally to the connection likelihood of two nodes. This paper proposes a model for link prediction, which is based on the node centrality of common neighbors. Three node centralities are discussed: degree, closeness and betweenness centrality. In our model, each common neighbor plays a different role to the node connection likelihood according to their centralities. Moreover, the weak-tie theory is considered for improving the prediction accuracy. Finally, extensive experiments on five real-world networks show that the proposed model can outperform the Common Neighbor (CN) algorithm and gives competitively good prediction of or even better than Adamic-Adar (AA) index and Resource Allocation (RA) index. Copyright c © EPLA, 2013 Introduction. – Given a snapshot of a network at time t, which new links or interactions among its members are likely to occur at time t′(t < t′)? We can formalize this question as the link prediction problem [1]. Link prediction is applicable to a variety of areas, such as protein-protein interaction (PPI) prediction [2], identifying spurious links [3], evaluation of network evolving mechanisms [4], e-commerce [5]. Zhou et al. [6] divided the link prediction algorithms into three categories: similaritybased algorithms, maximum-likelihood methods and probabilistic models. The similarity-based algorithms are the most used and they include node similarity and structural similarity. This paper will focus on node similarity algorithms. Node similarity link prediction algorithms rely on the low complexity, low time consumption and good prediction accuracy, which become one of the most applied link prediction approaches. Among which, Common Neighbor (CN) [7] is the most widely used node-similarity-based algorithm. The basic assumption is that two nodes x and y are more likely to have a link if they have many common neighbors. CN only considers the number of common neighbors. Further, many variants [8–10] of CN are proposed by taking the degrees of nodes x and y into account. Therein, the Preferential Attachment (PA) index [4] is suitable for the prediction of scalefree networks, where the probability that a new link is connected to the nodes x and y is proportional to the degrees kx and ky. Furthermore Adamic-Adar [11] and Zhou et al. [12] improved the CN by restraining the contributions of large-degree common nodes. They further improved the prediction accuracy. Most of the traditional approaches consider only the degree of each common neighbor of two nodes. They can improve the prediction accuracy, but, the improving is limited, because the node degree cannot reflect the significance of the node completely. Murata and Moriyasu [13] gave a weighted-common-neighbors approach. This paper assumed that proximities between nodes could be estimated better by using both graph proximity measures and the weights of existing links in a social network. It proposed a weighted graph proximity measures and new scores that took weights of links into account. Liu et al. [14] proposed a local näıve Bayes (LNB) model for link prediction in complex networks. In this model, different common neighbors will play different roles and give