Machine learning for link prediction in complex networks

Abstract

Nowdays, networks are omnipresent. The study and understanding of these networks become a greater need. The purpose of this work, is to investigate link prediction task in complex networks using Machine learning techniques. In fact, we propose two approaches to perform link prediction: supervised and unsupervised one. In both techniques a link or a pair of nodes is characterized by several features based on network topology-based metrics. In addition, we investigate many combined features. Concerning the supervised approach, we investigate the KNN and decision tree methods to build the link prediction models. While in the unsupervised approach, we rely on ranking strategy. An experimental study is performed on real networks. The results show that the supervised approach using gathered features reaches good performances with 84% f-measure.