A0879
Title: Calibrating covariates within product partition models, with an application to stochastic block models
Authors: Sirio Legramanti - University of Bergamo (Italy) [presenting]
Raffaele Argiento - Università degli Studi di Bergamo (Italy)
Valentina Ghidini - Euler Institute (Switzerland)
Abstract: Product partition models (PPM) represent a flexible framework for Bayesian nonparametric clustering. Their factorized structure facilitates the incorporation of individual covariates, giving rise to variants like the well-known PPM with covariates (PPMx) and spatial PPM (sPPM), with the latter being specialized to spatial covariates. Besides incorporating covariates into the clustering process, it is paramount to calibrate their influence on the obtained partition. A framework is proposed for weighting the influence of covariates within PPMs, and findings are illustrated on stochastic block models. The latter are models for clustering network nodes based on the network adjacency matrix. Such an application is further motivated by the fact that network data are often accompanied by node covariates. For example, in the real-data application to the public transportation network of Bergamo province (Italy), each network node has a spatial location, and one may aim for clusters that are as spatially cohesive as possible, while still reflecting the network structure. In fact, the obtained clusters may be used to inform policymaking in public transport, and it may be preferable that such policies are uniform over neighboring areas.
