Structure and Dynamics of Collaborative Knowledge Networks

A major means to encode and share scientific knowledge are publications, which cite each other and which are authored by one or more scientists. Citation networks of publications are commonly used to proxy the structure of scientific knowledge. Coauthorship networks are used to represent the social network between collaborating scientists. Yet, these two networks are rarely considered together even though they are interconnected. The multilayer collaborative knowledge network that results from combining the two allows us to study how the social relations among authors affect the structure and dynamics of the citation layer. To address this issue, we apply network theory. In the first part, we analyse the structure of collaborative knowledge networks. Our goal is to study dyadic interactions between individual pairs of authors in the context of the whole network. The ability to perform such a study will allow investigating individual citation behaviours of authors, as well as their deviations from community standards. For this, we develop a novel statistical method to extract how much authors' citations to each other deviate from a certain expectation. It builds on three methodological contributions. The first one is a flexible probabilistic model for complex networks that can encode heterogeneity in dyadic interactions. The second one is a procedure to formulate statistical null models for networks that respect temporal ordering of nodes and community structures. The third contribution is a new nonparametric probabilistic measure to quantify the deviation of an observed value from a distribution. With this method at hand, we present the deviations of authors' citations from the expectation formed based on the behaviour of the community at large. We also show how to use these deviations to highlight the intricate sub-community structures within the larger communities. In the second part, we study the evolution of collaborative knowledge networks. We show that the often neglected social layer has a significant effect on the citation layer. Particularly, we find that the overall likelihood of a publication to be cited scales with the number of previous publications by its authors, as well as with the number of their previous collaborators. To obtain this finding, we develop a method to fit and compare probabilistic growth models of multilayer networks. We further look into how the citations are distributed over time for a given publication and we find that citations arrive faster for the authors with more collaborators and more publications. The scientific contribution of this thesis is twofold. First, we develop novel statistical methods to study evolving multilayer complex networks. These methods can be applied in various fields. Second, we apply these methods to study citation and collaboration networks from the unified viewpoint of a multilayer network, which leads us to findings that could not be reached by merely considering the two layers in isolation