A1379
Title: Laplace approximations for Gaussian process and mixed effects quantile regression
Authors: Andrea Thomas Nava - Hochschule Luzern (Switzerland) [presenting]
Fabio Sigrist - ETH Zurich (Switzerland)
Abstract: The aim is to propose novel Laplace approximations for quantile regression with Gaussian process and mixed effects models that address limitations of standard Laplace approximations by replacing the Hessian with data-adaptive approximations. For large-scale applications with Gaussian processes, Vecchia approximations are used to enable scalable inference. The quality of the methods concerning posterior and marginal likelihood approximations is evaluated on both simulated and real-world datasets, benchmarking against existing state-of-the-art approaches for quantile regression with Gaussian processes and grouped random effect models. The methods are found to be computationally more efficient and stable while often providing more accurate posterior predictive distributions and hyperparameter estimates.
