Deep learning approaches for link prediction in social network services

Abstract

With the fast development of online Social Network Services(SNS), social members get large amounts of interactions which can be presented as links with values. The link prediction problem is to estimate the values of unknown links by the known links' information. In this paper, based on deep learning approaches, methods for link prediction are proposed. Firstly, an unsupervised method that can works well with little samples is introduced. Secondly, we propose a feature representation method, and the represented features perform better than original ones for link prediction. Thirdly, based on Restricted Boltzmann Machine (RBM) that present the joint distribution of link samples and their values, we propose a method for link prediction. By the experiments' results, our method can predict links' values with high accuracy for data from SNS websites. © Springer-Verlag 2013.