Generalizing link prediction: Collaboration at the University of Antwerp as a case study



The link prediction (LP) approach tries to predict links in an unknown network on the basis of a known network. It is argued that LP evaluation can be treated analogous to Information Retrieval evaluation. This characterization entails three generalizations of LP: both appearing and disappearing links can be predicted, LP is not necessarily time-based, and LP is complementary to anomalous link and gap discovery. Multi-input LP tries to increase precision and recall by having more than one known network as input. These concepts are applied to an informetric case study of collaboration at the University of Antwerp. Performance of different prediction methods is discussed. Furthermore, we establish a small but positive influence for multi-input LP.