Haavard Rue: "Group cross-validation and variational corrections in R-INLA"

Data i lloc: Divendres 19 de maig de 2023 · 12:00 h · Saló de Graus de la Facultat de Ciències Matemàtiques

Resum: In this talk I will discuss two recent developments in the R-INLA project, which aim to provide approximate Bayesian inference based in the INLA approach for the class of latent Gaussian models.

I would first discuss is our approach to correcting marginal density estimates by making use of the variational form of Bayes theorem [Zellner 1988]. We aim to correct the mean, variance and finally, the skewness; each is increasingly more involved. The overall design  riteria for doing these corrections, is to maintain good computational scalability with respect to model size and data size.

Secondly, I will discuss our new take on cross-validation (CV), an approach which is justified using independence-like assumptions. With dependent data, then leave-one-out CV make less sense, as we are evaluating interpolation properties rather than prediction properties. We can adapt the CV idea to dependent data, by removing a set of ``near-by'' data-points, before predicting, but the issue is then how to do this in practice, which is less evident for more involved models. I will discuss our approach which also can automatically can select appropriate groups of data, named group-CV.