Inférence et réseaux complexes
|Advisor:||Dubé, Louis J.; Desrosiers, Patrick|
|Abstract:||Modern science is often concerned with complex objects of inquiry: intricate webs of social interactions, pandemics, power grids, ecological niches under climatological pressure, etc. When the goal is to gain insights into the function and mechanism of these complex systems, a possible approach is to map their structure using a collection of nodes (the parts of the systems) connected by edges (their interactions). The resulting complex networks capture the structural essence of these systems. Years of successes show that the network abstraction often suffices to understand a plethora of complex phenomena. It can be argued that a principled and rigorous approach to data analysis is chief among the challenges faced by network science today. With this in mind, the goal of this thesis is to tackle a number of important problems at the intersection of network science and statistical inference, of two types: The problems of estimations and the testing of hypotheses. Most of the thesis is devoted to estimation problems. We begin with a thorough analysis of a well-known generative model (the stochastic block model), introduced 40 years ago to identify patterns and regularities in the structure of real networks. The main original con- tributions of this part are (a) the unification of the majority of known regularity detection methods under the stochastic block model, and (b) a thorough characterization of its con- sistency in the finite-size regime. Together, these two contributions put regularity detection methods on firmer statistical foundations. We then turn to a completely different estimation problem: The reconstruction of the past of complex networks, from a single snapshot. The unifying theme is our statistical treatment of this problem, again based on generative model- ing. Our major results are: the inference framework itself; an efficient history reconstruction method; and the discovery of a phase transition in the recoverability of history, driven by inequalities (the more unequal, the harder the reconstruction problem). We conclude with a short section, where we investigate hypothesis testing in complex sys- tems. This epilogue is framed in the broader mathematical context of simplicial complexes, a natural generalization of complex networks. We obtain a random model for these objects, and the associated efficient sampling algorithm. We finish by showing how these tools can be used to test hypotheses about the structure of real systems, using their homology groups.|
|Document Type:||Thèse de doctorat|
|Open Access Date:||19 October 2018|
|Collection:||Thèses et mémoires|
All documents in CorpusUL are protected by Copyright Act of Canada.