Predictive Network Representation Learning for Link Prediction

摘要

In this paper, we propose a predictive network representation learningrn(PNRL) model to solve the structural link prediction problem.rnThe proposed model defines two learning objectives, i.e., observedrnstructure preservation and hidden link prediction. To integrate therntwo objectives in a unified model, we develop an eu001dective samplingrnstrategy to select certain edges in a given network as assumed hiddenrnlinks and regard the rest network structure as observed whenrntraining the model. By jointly optimizing the two objectives, thernmodel can not only enhance the predictive ability of node representationsrnbut also learn additional link prediction knowledge inrnthe representation space. Experiments on four real-world datasetsrndemonstrate the superiority of the proposed model over the otherrnpopular and state-of-the-art approaches.