Latent Gaussian Processes for Distribution Estimation of Multivariate Categorical Data.

Multivariate categorical data occur in many applications of machine learning. One of the main difficulties with these vectors of categorical variables is sparsity. The number of possible observations grows exponentially with vector length, but dataset diversity might be poor in comparison. Recent models have gained significant improvement in supervised tasks with this data. These models embed observations in a continuous space to capture similarities between them. Building on these ideas and recent developments in sampling based variational inference we propose a Bayesian model for the unsupervised task of distribution estimation of multivariate categorical data.

Date:
Speakers:
Yarin Gal
Affiliation:
Cambridge University