Sosyal ağlarda bağlantı tahmini

Link prediction is used to forecast link evolution over time in networks. It has been used in several areas such as bioinformatics, online recommendation systems, e-commerce sites, collaboration networks and social networks. Predicting user behavior has become crucial with the expansion of multiuser online systems. This study aims to provide an insight to performance characteristics, both in terms of effectiveness and efficiency, for several link prediction methods. Four fundamental link prediction methods (i.e., common neighborhood, Adamic-Adar, preferential attachment, and Jaccard coefficient) that have been reported in the literature, and a novel metric have been evaluated. The proposed metric makes predictions on the premise that a newly joined member tends to make connections with available nodes that are popular amongst the network. Real-life data sets obtained from the Stanford Large Network Dataset Collection. Common neighborhood, Adamic-Adar and preferential attachment metrics provided more successful results than the others in all networks. In terms of running time, preferential attachment, common neighborhood and the novel metric of this study are the fast-running ones. The highest F1-score is 0.12 in the email-Eu-core and Reddit networks achieved by the Adamic-Adar metric. This study presents and discusses the performance of several link prediction methods on temporal networks. It provides some insights for practical usage of link prediction metrics. Keywords: Link prediction, structure based metrics, temporal network