A0510
Title: Dependent stochastic block models for sequences of directed networks with application to causes of death co-occurrences
Authors: Giovanni Romano - Bocconi University (Italy) [presenting]
Cristian Castiglione - Bocconi University (Italy)
Daniele Durante - Bocconi University (Italy)
Abstract: A new Bayesian model is developed for inferring changes in the stochastic block structures within a sequence of weighted and directed networks indexed by a predictor. This model, named dynamic stochastic block model (dSBM), is originally motivated by the demographic problem of learning age-specific group structures in directed networks encoding co-occurrences among underlying and multiple death causes for different age groups in a population. To this goal, state-of-the-art stochastic block models are substantially generalized to account for (i) edges with categorical weights, (ii) two separate node partitions for their different roles in a directed network, i.e., sender/receiver, and (iii) sequences of networks indexed by an ordered predictor, such as age or time. Results in the causes-of-death networks analysis unveil interesting and yet-unexplored patterns in the composition, evolution, and interactions of causes-of-death clusters, with relevant analytic and policy implications. The novel formulation we propose provides a powerful model to infer non-trivial changes in grouping structures governing sequences of dynamic directed networks, beyond the demographic application.
