Link Prediction via Second Order Cone Programming

Abstract

Link Prediction has emerged as an important problem with the recent interest in studying large scale social graphs. User interactions on social networks can be represented as signed directed graphs where the links represent nature of their relation. Positive links correspond to trust/friendship among nodes. Negative links typically map to distrust or antagonism among the graph nodes and are useful in analyzing social graphs when coupled along with positive links. In this paper, we study the prediction of positive and negative links in a large scale graph. We propose a classification-afterclustering approach and design maximum margin classifiers which can be formulated as standard Second Order Cone Programs. Our proposed approach is immune to class imbalance and scales to large graphs very well. We apply this to various scenarios of link prediction problems of separating between negative links and positive links, globally and within a neighborhood. Empirical evaluations is reported on a real world social graph.