A0798
Title: Time-dependent stochastic block models with application to causes of death networks
Authors: Cristian Castiglione - Bocconi University (Italy) [presenting]
Abstract: Discovering latent dependence structures over the nodes of a dynamic network is a difficult challenge which is of increasing importance in many applied fields. The interest is motivated by a demographic analysis of the complex interaction between underlying and multiple causes of death in a population observed on a fine age grid. To unveil non-trivial grouping structures between causes of death, a flexible stochastic block model is proposed for dynamic directed networks, which is able to learn asymmetric node partitions having common connectivity patterns. To flexibly account for the time evolution of the node grouping structure, a time-dependent random partition process is relied on, which permits the learning of sequences of partitions with a high level of persistence over time. The initial distribution of the sequence is then specified according to a Gibbs-type prior. This choice encompasses several routinely used prior distributions for Bayesian clustering, including fixed, random and infinite number of possible groups. An additional benefit of such a specification is to facilitate the inclusion of non-dynamic node-specific attributes in the model, which, in the application, permits the information of the partition sequence with external medical information on the causes of death. Alongside, an efficient algorithm to sample from the posterior distribution of the considered model is discussed.
