Link prediction in weighted complex networks using machine learning methods and similarity metrics

The extensive applicability of social network analysis in capturing social interactions has attracted substantial interest from scholars. One crucial aspect of social network analysis is link prediction, which pertains to estimating the likelihood of a connection between two unconnected nodes within a network. Numerous studies have been conducted to address this problem, and among the existing approaches, similarity-based methods have demonstrated notable efficiency. These methods consider the network’s structure and employ the number of shared neighbors between two nodes as a fundamental measure for establishing structural similarity. When nodes exhibit high structural similarity, it suggests a higher likelihood of a link between them. In this study, two tasks of link prediction are explored. Firstly, the prediction of links in weighted complex networks is investigated using similarity-based metrics. Secondly, link prediction is formulated as a classification task to showcase the efficacy of similarity-based metrics when combined with machine learning models. To evaluate the performance of link prediction metrics, a collection of network-weighted datasets is utilized, and the area under the curve is employed as an evaluation measure. Finally, various supervised learning algorithms are assessed based on their prediction accuracy.