Probabilistic Latent Tensor Factorization

We develop a probabilistic modeling framework for multiway arrays. Our framework exploits the link between graphical models and tensor factorization models and it can realize any arbitrary tensor factorization structure, besides many popular models such as CP or TUCKER models with Euclidean error and for nonnegativity with KL error. The probabilistic framework enables us to develop a model selection methodology based on variational Bayes. 1 The Model : Probabilistic Latent Tensor Factorization (PLTF) We propose a unifying framework for full Bayesian inference in which any arbitrary tensor factorization structure for Euclidean and KL costs can be realized. By making use of the duality between exponential families and Bregman divergences, here we cast the Tensor Factorization problem as inference problem in the probabilistic Graphical Models with Gaussian or Poisson components. This way the tensor factorisation reduces to a parameter estimation problem. The associated inference algorith...

• Type: Conference
• Year: 2010
• Where: ICA
• Authors: Y. Kenan Yilmaz, A. Taylan Cemgil
• Keywords: Arbitrary Tensor Factorization, ICA 2010, Signal Processing, Tensor Factorization, Tensor Factorization Structure