Projection predictive input variable selection for Gaussian process models
Juho Piironen, Aki Vehtari · arXiv (Cornell University) · 2015
We discuss the problem of selecting input variables for a Gaussian process (GP) model based on their predictive relevancy. We propose a new projection framework for training and comparing models with subsets of the candidate variables by utilizing the information in the full model, fitted with all the inputs. Our results on synthetic and real world datasets indicate that this approach improves the assessment of input relevancies and the fit of the actually selected submodels by utilizing the uncertainties contained in the full model. In particular, we demonstrate that our method generally outperforms the automatic relevance determination (ARD) and the use of marginal probabilities in determining the input relevancies.