Link prediction using betweenness centrality and graph neural networks

Social networks are widely considered as the most important tool to connect people. The last century saw a massive increase in the number of links between users. Many nodes and/or edges are included or removed repeatedly as time passes. In order to understand the patterns of relations among people or organizations, social network analysis (SNA) sheds light on users and links dynamics. Link prediction (LP) is one of the most important research areas of SNA. The main objective of link prediction is to determine whether two nodes will form a link or not in the future. LP uses similarity-based methods such as common neighbors, resource allocation, and Adamic–Adar metrics to forecast potential connections from the current state of networks. Although using the similarity-based methods is highly time efficient, the measures still suffer from low accuracy. The main focus of this paper is to address this drawback by defining a new metric called LSBC that uses the combination of a similarity metric and the betweenness centrality which defines the node’s power over the entire network. The method was illustrated with nine datasets from different types of networks. Experiments show that LSBC captures the similarity of a pair of nodes accurately and surpasses all the state-of-the-arts methods. Furthermore, we use neural network model to address the link prediction as a classification task. The results show that the adding LSBC as additional feature increases the accuracy and reduces the cross-entropy loss.