Abstract:
This thesis contains three essays in learning representation of complex networks, the first two of which develop new methods and the third utilizes these methods in a real-world application. The first essay provides methods for extracting underlying signed network backbones from intrinsically dense weighted networks. Utilizing a null model based on statistical techniques, we propose significance and vigor filters that enable inferring edge signs and weights. Empirical analysis on four real-world networks reveals that the proposed filters extract meaningful and sparse signed backbones that exhibit characteristics typically associated with signed networks while respecting the multiscale nature of the network. The second essay deals with the misalignment problem in dynamic representation learning. We provide the first formal definitions of alignment and stability, propose novel metrics for measuring them, and show their suitability through a set of synthetic and real-world experiments. We show that, by ensuring alignment, the performance of dynamic network inference tasks improves by a remarkable amount. The third essay applies the novel methods developed in the first two essays as well as other methods from the network analysis literature to investigate the structure and dynamics of internal migration in Turkey. In addition to providing unique and specific insights, we find that most migration links are geographically bounded with exceptions of cities with large economic activity, migration takes place in well-defined routes, counter-streams develop for major migration streams, and the migration system is largely stable over time; which are generally in line with classical migration laws.
