A1153
Title: Enriched stochastic block models
Authors: Louise Alamichel - Bocconi University (Italy) [presenting]
Daniele Durante - Bocconi University (Italy)
Tommaso Rigon - University of Milano-Bicocca (Italy)
Francesco Gaffi - University of Notre Dame (United States)
Abstract: Stochastic block models learn group structures among nodes sharing similar connectivity patterns. Recent extensions have considered scenarios where multiple networks sharing the same nodes need to be analyzed jointly. State-of-the-art formulations typically assume that a unique partition of the nodes governs the clustering across all networks. In practice, the node partition describing block structures in one network may differ from that underlying another network, with the two often linked by refinements for specific purposes. For example, for two networks, the grouping of nodes in one layer may fragment into more detailed communities in another layer, reflecting different but related structures. To capture these architectures, we develop a novel enriched stochastic block model relying on two clustering structures, one for each network, and linked through a joint enriched Bayesian nonparametric prior. This construction induces a dependence across partitions, while retaining flexibility to model both shared and network-specific patterns. Inference under the proposed model is carried out via a collapsed Gibbs sampler on the cut posterior. Preliminary results on simulated data and in a study of summits co-attendances within a complex Mafia organization showcase the strengths of the proposed formulation, along with its ability to incorporate and learn relevant structures in networks.
