ED Mathématiques et Informatique
Solving combinatorial optimization problems by hybrid methods combining machine learning techniques with existing techniques
by Fulin YAN (IMB - Institut de Mathématiques de Bordeaux)
The defense will take place at 10h00 - Salle de conférences de l'IMB 351, cours de la Libération CS 10004 - Bâtiment A33 33405 Talence CEDEX
in front of the jury composed of
- Francois CLAUTIAUX - Professeur - Université de Bordeaux - Directeur de these
- Nicolas JOZEFOWIEZ - Professeur - Université de Lorraine - Rapporteur
- Vincent T'KINDT - Professeur - Université de Tours - Rapporteur
- Christine SOLNON - Professeure - INSA de Lyon - Examinateur
- Thibault PRUNET - Maître de conférences - Université de Bordeaux - Examinateur
In this thesis, we focus on combinatorial optimization problems that can be formulated as arc flow problems derived from dynamic programming. We investigate the potential of hybrid methods that integrate machine learning to guide heuristic algorithms for solving such formulations. First, we provide background and a literature review on arc flow formulations and machine learning–based hybrid approaches for combinatorial optimization. Second, we study the use of machine learning as a scoring function within a beam search algorithm to find high-quality solutions for a resource-constrained shortest path problem in very large acyclic graphs with a large number of resource constraints. Third, we investigate the use of machine learning to sparsify the underlying graph of arc flow formulations to reduce problem size while enabling faster computation of high-quality solutions. To validate our approaches, we present computational results on several problem classes and compare them with the same methods without incorporating machine learning. The results demonstrate the effectiveness of both machine learning–augmented beam search and machine learning–guided graph sparsification, showing significant improvements in solution quality on average across multiple problems. Finally, we conclude with a discussion of the results and outline perspectives for future work.
ED Entreprise Economie Société
gender disparities in academic careers: productivity, collaboration Networks, and research agendas
by Amal BOUGHNIM (BSE - Bordeaux sciences économiques)
The defense will take place at 9h00 - Salle des thèses Salle des thèses Campus de Pessac Avenue Léon Duguit 33600 Pessac
in front of the jury composed of
- Pascale ROUX - Professeure des universités - Université de Bordeaux - Directeur de these
- Reinhilde VEUGELERS - Professor - KU-Leuven, Department of Management, Strategy and Innovation - Rapporteur
- Nicolas CARAYOL - Professeur des universités - Université de Bordeaux - CoDirecteur de these
- Hanna HOTTENROTT - Professor - Technical University Munich (TUM), School of Management - Rapporteur
- Emeric HENRY - Professeur des universités - Science Po Paris, Département d'économie - Examinateur
- José DE SOUSA - Professeur des universités - Université of Paris Panthéon-Assas - Examinateur
Gender disparities in science represent not only an issue of fairness but also a constraint on knowledge production. In this dissertation, I investigate these disparities in French academic research along three dimensions: scientific productivity, collaboration networks, and thematic specialization. I rely on a dataset covering 93,304 academics and their publications over 60 years, up to 2023. In the first chapter, I examine the gender productivity gap across the career cycle. The results reveal a U-shaped trajectory: the gap is already present at age 30 (~17%), widens to ~31% at age 42, then partially narrows without ever closing. This pattern, robust across cohorts and disciplines, reflects distinct dynamics: men's productivity rises faster, plateaus higher, then declines after age 50, while women's experiences a late-career resurgence. Life sciences are a notable exception, with an ever-widening gap. A coarsened exact matching shows that initial selection accounts for one third of the differences but does not affect the U-shaped profile. Parenthood analysis at a large French university reveals no gender-differentiated penalty. However, ANR data (2005–2018) show that women are 15% less likely to apply for grants, a difference 2.5 times larger in life sciences. Finally, a triple-difference approach shows that environments combining teaching and research disproportionately reduce women's output. In the second chapter, I study coauthorship networks. Progressive specifications controlling for career length and productivity reveal that women's apparent network size disadvantage is largely associated with career length. Controlling for total productivity reverses the gap entirely: women achieving comparable output do so with larger, more diverse, and higher-quality networks, suggesting greater collaborative capital is needed to reach parity. Exploiting the LabEx funding program confirms these results: funding generates a substantial expansion of women's networks through new ties formed outside the funded cluster, while men's gains remain concentrated within it. Women improve their structural position by gaining bridging weak ties, while men experience the opposite pattern. These findings are consistent with binding constraints on access to collaborative opportunities rather than differential preferences. The third chapter analyses research orientation among 3,950 French mathematicians. The gender gap is one of volume, not quality: 17% fewer papers, but comparable citations and journal quality. Women occupy a systematically different position in the mathematical landscape, such that gender diversity is itself a source of research diversity. Their concentration in higher-output applied subfields masks the within-subfield gap, which rises from 17% to 25% at comparable subfield positioning. Their higher cross-disciplinary engagement, particularly toward life and medical sciences where they are twice as likely to publish, partially offsets this gap. The net effect of controlling for lagged within-field specialization and cross-disciplinary engagement is to widen the productivity gap to 27%, indicating that women's research agendas function as a source of resilience. These patterns vary by initial productivity: the least productive women diversify within mathematics, moderately productive ones reach toward other disciplines, and the most productive converge toward fundamental subfields.