Description du sujet de thèse

Since the invention of the printing press, media influence on social opinion has sparked debate at every technological leap, becoming a battleground for political and economic forces seeking control. Online social media, now widely used as primary information channels and tools for interaction, are no exception. What sets them apart is the unprecedented speed of communication and sheer volume of exchanged data, creating a vast public arena with evolving rules of virtual interaction. These rules emerge from complex dynamics, blending individual motivations with rapidly advancing technology and economic constraints
This PhD proposal is related to the COBCOD ANR project (2025-2029) based on the hypothesis that this new landscape not only shapes worldviews but also influences how we perceive and construct our social identities. Algorithms that govern content exposure and suggest social connections play a key role, often operating behind closed doors due to industrial property rights. Evidence suggests they amplify cognitive biases such as negativity and confirmation bias, leading to collective phenomena like herding, polarization, and echo chambers. Moreover, agent-based models indicate that even without algorithmic filtering, confirmation bias alone can drive opinion polarization.
In light of these findings, one of the core hypothesis of this project is that global effects seen in online social networks are partly driven by well-known cognitive biases. Due to the unique features of online platforms—such as high diffusion speed, selective exposure, message size limitations, and potential algorithmic biases—these biases significantly shape interaction dynamics within the network, leading to phenomena not observed offline.
COBCOD's primary objective is to better understand how cognitive biases are amplified and potentially altered through repeated online interactions. Specifically, we seek to identify the mechanisms driving the emergence of dynamic patterns unique to online environments and explore how these collective dynamics, in turn, influence individuals.
Our second objective is to investigate strategies for mitigating these effects and test them through models and experiments. We will explore two approaches: examining how users behave when informed about cognitive biases and their effects, and assessing the impact of modifications to the user interface during online interactions.
The project will conduct a multi-disciplinary approach to achieve these objectives which includes individual-based modeling, complex computer simulations using both standard and generative agent-based models (ABM and G-ABM) to study the effects of biased agents in online networks, theoretical analysis of emergent phenomena, large-scale social network data analysis, online experiments with volunteer users, and fully controlled laboratory experiments.
The program is structured around team synergy, where individual-scale experiments inform theoretical models of interacting agents, with their properties validated by those experiments. The resulting dynamics will be compared with phenomenological observations, and the reverse approach will also be explored to identify potential feedback loops.
The PhD position will be centered on the data-mining and modeling part : reconstructing real opinion dynamics from large scale empirical data and modeling opinion dynamics calibrated on those reconstructions.
More precisely, the PhD will be focused on the following research directions :
• How can we develop metric spaces to measure individual opinions and their evolution, reconstruct global opinion dynamics from large-scale Twitter data, and identify shifts within social groups along with their potential causes. This research will be based on longitudinal, already captured, Twitter datasets from CYU and CNRS/ISC-PIF (project partners), covering US and French politics (since 2016), climate change debates (since 2016), and the COVID-19 pandemic (since 2020). With over six years of data and up to half a billion tweets per dataset, it offers a unique opportunity to track opinion changes across large social groups amid major crises (e.g., gilets jaunes, climate, pandemics).
• How can we develop methods deriving numeric scales and values from textual
messages and other interactions (retweets, likes etc.), specifically for each bias: emotional valence for negativity bias, opinion extremity (between extremely favorable to extremely unfavorable) for confirmation bias, feelings towards others for in-group bias.
How to produce indicator and visualization to qualify quantitatively and
qualitatively these dynamics.

Overall, the PhD will address the question of identifying the cause for large scale opinion shifts.

Contexte de travail

The doctoral student will conduct their thesis under the supervision of David Chavalarias (HDR) and will work at ISC-PIF within the computational social sciences team, in collaboration with other COBCOD project partners. The workplace will be located at the Institut des Systèmes Complexes de Paris Île-de-France, 113 rue Nationale 75013. He/she will be enrolled at the EHESS Doctoral School 'Sciences de la Société' (Social Sciences).

The Institut des Systèmes Complexes de Paris Île-de-France (ISC-PIF, http://iscpif.fr) is a CNRS Research Support Unit with around twenty staff members (tenured and contract employees). Combining the efforts of 12 EPSTs in the Paris region around interdisciplinary research on complex systems, ISC-PIF functions as a shared resource centre and project hub. It provides access to resources that are beyond the reach of traditional disciplinary approaches by pooling skills and resources (databases, computing resources, shared platforms, algorithms, etc.). The ISC-PIF's technological platforms are at the heart of several dynamic international user communities.

The ISC-PIF hosts numerous shared services in the field of data analysis (particularly concerning the analysis of large text corpora) and high-performance computing. These services are accessible remotely via ad hoc interfaces; they use and generate large amounts of data. The PhD student will work in particular on the databases of the Politoscope and Climatoscope platforms.