| The aim of this dissertation is to develop methods to predict new or missing interactions based on information in the form of a multi-relational network. It is based on two case studies of collaboration between researchers: collaboration at a mid-sized institute (the University of Antwerp) and collaboration within a research specialty (informetrics). The case studies can be interpreted as networks, such that each node represents an author and each link represents co-authorship.;After providing a general overview of methods and techniques in (social) network analysis, we introduce Q-measures, indicators that characterise the extent to which a node plays a bridging role between groups in the network.;When viewed dynamically (over time), networks evolve. The link prediction problem is the question to what extent changes in link structure can be predicted. We introduce a general framework for link prediction, consisting of five steps. Empirical examination reveals that path-based predictors offer the best overall performance and that weighted predictors typically perform worse than their unweighted counterparts. Furthermore, we show that network reconstruction (reconstructing a damaged network) can be used to assess a method's applicability to link prediction.;In a following step, we broaden our approach to multi-relational (or 'semantic') networks, i.e. networks that consist of links and nodes of different types. We discuss how these concepts are related to semantics and provide an introduction to Semantic Web technologies. This raises the question whether an RDF-based knowledge representation can be created that is suitable for informetric research. We discuss the main difficulties and argue that existing ontologies provide the bulk of the vocabulary needed for informetrics. Where necessary, we suggest additions to the existing infrastructure, e.g. to capture term usage.;We show that the bipartite author–paper network is a better training network than the derived co-authorship network for specific predictors. Next, we substitute the author–paper network with an author–term network. Finally, we introduce the association network model, a general model that can be used to identify and study patterns of interest in a multi-relational network. This model's main value lies in the context of recommendation. |
