A0346
Title: Dynamic network models with time-varying nodes
Authors: Luca Gherardini - University of Florence (Italy) [presenting]
Abstract: Networks are used to represent complex data structures that arise in different fields of science and are often not static objects, as they can evolve over time. The main approaches in the literature belong to the broad class of latent variable models, including latent space models and stochastic block models. However, most methods were developed to model the dynamic behaviour of edges without considering that the network's topology may vary over time. It is shown that ignoring this new source of complexity can lead to non-negligible bias in the parameter estimates since the model cannot discriminate structural zeros due to the topology of the network from those related to the absence of an edge between a pair of nodes. A class of zero-inflated Bernoulli model is proposed that embeds the time-varying node probabilities into the edge process, with the aim of modelling the entire dynamic network over time. A fully latent approach is also proposed for the nodes process, providing a probabilistic characterization for any network model (static and dynamic) that explicitly assumes a two-layer hierarchy, one for nodes and one for edges, regardless of the modelling choice. The inference approach for this class of models is developed within the Bayesian paradigm and relies on a Gibbs sampling algorithm with a Polya-gamma data augmentation scheme. The performance of the approach is explored through a simulation study.
