Empirical Evaluation of Link Prediction Methods in Social Networks

Abstract

Link prediction in social network has attracted increasing attention from a broad range of communities. In this study, we examine the predictive performance and time-efficiency of two group of methods for this problem. The first group consists of similarity metrics, including Jaccard Coefficient (JC), Adamic-Adar Coefficient (AA), Preferential Attachment (PA) and Personalized PageRank (PPR). The second group comprises embedding methods, including Laplacian Eigenmaps (LE), Node2Vec and Variational Graph Auto-Encoders (VGAE). All methods were evaluated extensively on Facebook EgoNets dataset. We observe that Node2Vec is the most efficient method in terms of training time and accuracy on many types of graph. Besides, we also give insights into the properties of these methods, which can be a basis for further research on this topic.