I will defend my HDR entitled ‘Combining Heterogeneous Information: Contributions to the Extraction and Analysis of Feature-Rich Complex Networks’ on Wednesday, June 16, 2021, at 9 a.m. The presentation will be in English and will take place via video conference at the link https://pod.univ-avignon.fr/live/colloque/
You are all cordially invited to attend.
Jury:
- Céline Rouveirol, Professor, Université Sorbonne Paris Nord, LIPN, President
- Éric Gaussier, Professor, Université Grenoble Alpes, LIG, Reviewer
- Jean-Loup Guillaume, Professor, Université de La Rochelle, L3i, Reviewer
- Maguelonne Teisseire, Research Director, INRAE, UMR Tetis, Reviewer
- Roger Guimerà, Professor, Universitat Rovira i Virgili, ICREA, Examiner
Abstract: The concept of Complex Network is generally used in the literature to refer to a graph representing a real-world complex system. This confers such graphs so-called non-trivial topological properties that distinguish them from regular and random graphs. Among them, the most widely known are small-worldness and scale-freeness, whose study marked the beginning of a new research domain now called Network Science, and aiming at studying complex networks. It is a multidisciplinary field that relies largely on a number of pre-existing domains, in particular graph theory, quantitative sociology, computer science, operations research, statistical physics, and of course complex systems. Network Science is mainly a data science, as its starting point is the modeling of real-world systems. As such, its emergence is due not only to the convergence of interdisciplinary efforts, but also to the availability of the resources required to build and study large and/or numerous complex networks: computational power and access to sizable datasets. Because of this fundamental reliance on data, information representation is a fundamental aspect of Network Science. The way the available data describing the considered system are included in the graph-based model is of the utmost importance. Yet, plain graphs are meant only to model one type of information: the presence or absence of relationships between the object constituting the system. To handle more diverse data, it is necessary to extend this framework, which leads us to the notion of Feature-Rich network that is at the core of this thesis. In this manuscript, I summarize the research that I conducted on the topic of feature-rich networks. In the first part, I focus on vertex-attributed networks. Chapter 2 deals with community detection, and presents a comparison of vertex modules detected based on graph structure vs. attributes, using a student activity dataset collected during a ground survey. Chapter 3 tackles two classification problems based on attributed graphs. The first is the detection of persons which are influential in real-life, based only on data describing their activity on an online medium. The second is the identification of abusive messages in online chats. The second part is dedicated to dynamic networks. Chapter 4 proposes two methods leveraging sequential pattern mining to characterize the dynamics of such networks at two distinct levels. The first targets the microscopic evolution of the network, whereas the second describes the network at a mesoscopic level. In Chapter 5, I present a method based on the segmentation of dynamic graphs, and aiming at generating video summaries of TV series. The third part covers three types of networks. Chapter 6 is dedicated to spatial networks, presenting two works revolving around the Straightness measure. Chapter 7 is about the partitioning of signed networks and the concept of structural balance. Chapter 8 deals with multiplex networks. I describe a vertex centrality measure relying on a model of opinion diffusion in multiplex networks, and a graph partitioning method allowing to identify several modular structures for a single multiplex network.