Predicting Links in Multi-relational Networks

Abstract

Most traditional methods of link prediction in networks deal with homogeneous networks, i.e., networks with a single type of nodes and a single type of relationships. However, most real-life systems modelled by networks comprise multiple types of entities undergoing multiple types of interactions among themselves. In most social settings, for example, the actors are often connected by multiple types of social ties at the same time (friendship, kinship, acquaintance, FB-friend, twitter-follower, LinkedIn-contact, etc.). The types of ties a node already has influences which other ties it will form in the future -- thus not only the existing link structure, but the variation in the link structure in terms of relationship types determines the \emph{target} as well as the \emph{type} of the new ties. In this paper we propose a novel method for link prediction in multi-modal and multi-relational networks. Our method is based on semantics of simple and compound relationships in the given network, i.e., a compound relationship represents a well-defined pattern of simple relationships between two typed nodes in the network. We test our method with Sageman's data set of the Salafi Jihad network, which is a heterogeneous network comprising of multiple relationship types.