Machine Learning in Montpellier, Theory & Practice – Mathilde Perez (Télécom Paris)
Real-world networks are rarely static: interactions accumulate over time, shaping who connects with whom and how influence spreads. This talk presents two works that model this temporal dimension explicitly, using temporal point processes (Hawkes processes) as a common underlying tool. First, SimHawNet introduces a generative model for continuous-time temporal networks, coupled with a simulation algorithm that reproduces the structural and temporal dynamics of real-world networks. Second, we study homophily, the tendency of similar individuals to connect, in evolving networks, and separate how much of it reflects genuine individual preference (« choice » homophily) from how much is amplified by the system itself, as algorithms repeatedly recommend and reinforce the same type of connections (« induced » homophily). We introduce an instantaneous bias measure that captures this induced component as it happens, revealing feedback effects that static, snapshot-based fairness metrics miss entirely.