A0693
Title: Bayesian co-clustering of ordinal data with informative censoring
Authors: Alice Giampino - University of Milano-Bicocca (Italy) [presenting]
Antonio Canale - University of Padua (Italy)
Bernardo Nipoti - University of Milan Bicocca (Italy)
Abstract: A novel Bayesian nonparametric model is proposed for co-clustering multivariate ordinal data. The ordinal nature of the data is addressed through a latent variable framework, while its large dimensionality is managed via a matrix factorization model. The flexibility of the approach stems from the use of two independent Dirichlet processes, which allow for inference on the number of clusters within the latent structure. Unlike most approaches available in the literature, the method treats censored observations as potentially informative rather than absent information, reflecting that missing information can be valuable in profiling individual preferences. Thanks to the conjugate specification of the model, which allows for the explicit derivation of the full conditional distributions, posterior inference is performed using a Gibbs sampling algorithm. The performance of the method is demonstrated using data on politician votes.
