Year: 2026
Type: preprint
Abstract: Relational event data record time-stamped interactions between actors. Dynamic network actor models represent each interaction as two separate decisions: when an actor becomes active and whom they select. Modelling these decisions separately cannot establish whether the most active actors are also the most frequently selected. We propose a latent space dynamic network actor model in which sender and receiver positions capture receiver choice, while a position-rate link relates latent positions to activity levels. The link uses two summaries invariant to latent position updates: latent popularity, measuring how much an actor is favoured by others, and excess self-affinity, capturing preference for similar actors beyond that explained by popularity. We adopt a Bayesian modular approach, estimating latent positions from receiver choices alone. Using e-mail data from a manufacturing company, we find that the most active employees are also the most frequently selected recipients. With multiple latent dimensions, preference for similar colleagues contributes less. Latent positions improve receiver-choice prediction, while the position-rate link captures differences in actor activity.
Source: arXiv (Cornell University)
Authors: Caimo, Alberto, Gollini, Isabella
Language: English
Subfield: Statistical and Nonlinear Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Keywords: latent space models, Bayesian estimation
Open Access status: green
OpenAlex ID: W7236866574
