Relational event mixture models

Relational event mixture models

Catherine Matias, Sorbonne Université, CNRS & Université Paris Cité.

17th of September 2026

Abstract: 

We propose a relational event mixture model that accounts for unobserved heterogeneity at the level of interaction intensities in continuous-time networks. Conditional on the latent classes, the interactions follow an inhomogeneous Poisson process with intensity described by a time-varying cluster-specific baseline as well as cluster-specific endogenous and exogenous effects. We develop a scalable inference strategy based on a variational expectation-maximization framework. Under a nested case-control sampling, the maximization step corresponds to a degenerate weighted logistic model. Thus, cluster-specific nonlinear smooth effects can be fitted efficiently also in the case of relational event mixture models. We select the number of latent groups by an integrated classification likelihood criterion that accounts for the smoothness of each cluster-specific hazard. The effectiveness of the method is demonstrated through a simulation study and an empirical illustration.

Joint work with Veronica Poda, Rebecca Dallabetta, Veronica Vinciotti, and Fanny Villers.