A0450
Title: Bayesian adaptive Tucker decompositions for tensor factorization
Authors: Federica Stolf - University of Padova (Italy) [presenting]
Antonio Canale - University of Padua (Italy)
Abstract: A novel Bayesian method is presented for low-rank tensor decomposition of multiway data with missing observations. An adaptive Tucker decomposition model (AdaTuck) is introduced that automatically infers the latent multi-rank in a principled manner using an increasing shrinkage prior. The proposed approach introduces local sparsity in the core tensor through a shrinkage prior, inducing rich and, at the same time, parsimonious dependency structures. Posterior inference proceeds via an efficient adaptive Gibbs sampler, allowing for straightforward missing data imputation and making it well-suited for high-dimensional datasets. Simulation studies and applications to complex spatiotemporal data provide compelling support for the new modeling class relative to existing tensor factorization methods.
