Extraction of information from large networks

Every day, millions of customers of mobile phone operators communicate via phone calls, SMS or MMS. These interactions can be represented by a large network, where nodes represent customers, and links are drawn between customers that have had a phone call or exchanged messages. This example is one of the numerous applications in which large networks, mathematical models of interactions between millions of items, are used to represent large datasets of networked systems. The study of the structure of such networks provides useful insights on their organization. In this work, we address three different topics in the extraction of information from large networks. In the first part of this work, we focus on geographical networks, i.e. networks where every node is associated to geographical coordinates. The availability of this information allows studying how geography influences the creation of links. In particular, the intensity of communication between nodes decreases as a power of the geographical distance that separates them. Second, we address the topic of networks where links change over time, called dynamical networks. In dynamical networks, new nodes enter or leave the network and the strength of their connections rise and wane during the observation period. A phenomenon that is poorly understood so far is the impact that time scales play on the emergence of different structural properties of dynamical networks. In this second part of the thesis, we show how the dynamics of data-driven networks are composed of a complex interaction of processes having each a different characteristic time scale. Finally, the last part of this work concerns the detection of communities in networks. Communities are groups of nodes that are densely connected to each other. We propose an improvement of actual multi-level community detection methods in which the intermediate results are used at subsequent steps to avoid unnecessary calculations.(FSA 3) -- UCL, 201