PhD defense of Gaspard ABEL

Gaspard ABEL

Structural and Dynamical Principles of Information Propagation in Socio-Computational Systems.

Abstract

Online social platforms and autonomous computational agents have transformed the production, transmission, and processing of information. In these socio-computational systems, information does not evolve independently of the entities that process it or the interactions through which it is exchanged. Instead, user behavior, network structure, learning dynamics, and information content are continuously coupled, giving rise to collective phenomena such as information cascades, competing contagions, political polarization, representational collapse, and cooperation.

This article-based thesis investigates these phenomena through a common conceptual perspective based on three interlinked pillars: Information, Individuals, and Interactions. The objective is not to provide a single model of all networked systems, but to project this perspective onto several complementary problems involving social networks, diffusion processes, neural networks, and multi-agent systems. The work combines tools from applied mathematics, statistical physics, network science, dynamical systems, and machine learning.

First, the thesis introduces MIC, a marked multidimensional Hawkes process that jointly models multiple information cascades and user activity. By explicitly representing cross-cascade and user--cascade dependencies, MIC improves the modeling and prediction of information diffusion while providing interpretable parameters for the analysis of socio-semantic network structure. This interpretability is complemented by MultiLayerNetViz, a lightweight Python toolkit for visualizing multilayer networks in two and three dimensions, which is applied to political and social media datasets.

Second, the thesis addresses the control of collective dynamics on graphs. The gLRIE strategy extends dynamic resource allocation to competing contagions evolving under a continuous-time SIS-like model. Its theoretical and empirical analysis shows how network-aware interventions can be used to favor a desired state. A related control perspective is applied to Graph Neural Networks, where bifurcation theory is used to characterize oversmoothing as convergence toward a stable homogeneous representation. Bifurcation-inducing activation functions and a bifurcation-aware initialization are then proposed to generate stable, non-homogeneous node representations.

Third, the thesis studies how structured representations and learning rules can shape neural and artificial agents. The Dual Criterion Curriculum Learning framework combines loss-based and density-based criteria to construct curricula for time-series forecasting. Finally, a Q-learning analysis of game-theoretic scenarios shows how asynchronous and heterogeneous update rates affect the emergence of cooperation, which provides insights on regulation perspectives for algorithmic pricing in economic markets.

Together, these contributions provide a unified viewpoint for studying how structured information, heterogeneous individuals, and their interactions shape the dynamics of complex socio-computational systems, and how these dynamics can be modeled, interpreted, and leveraged for learning.

PhD supervisors  

• Julien RANDON-FURLING
• Jean-Pierre NADAL
• Argyris KALOGERATOS

Jury members

• Amandine VéBER, Directrice de recherche, Université Paris-Cité, Rapporteur
• Bruno LOUREIRO, Chargé de recherche, ENS Paris, Rapporteur
• Eva LöCHERBACH, Professeure des universités, École Polytechnique, Examinateur
• Fabien TARISSAN, Directeur de recherche, ENS Paris-Saclay, Examinateur
• Alfredo BRAUNSTEIN, Associate Professor, Politecnico di Torino, Examinateur

Keywords

Complex networks, Social networks, Opinion dynamics, Machine Learning, Representation Learning

Publications