useR! 2024

The qgam R package is an extension of the mgcv package, offering methods for building and fitting quantile additive models (QGAMs), which do not make any parametric assumption on the distribution of the response variable. While QGAMs make fewer assumption than standard GAMs, they are slower to fit due to the cost of selecting the so-called “learning-rate”. The longer fitting time is particularly problematic when handling large data sets and complex models. This talk focuses on the development of new Big Data methods for QGAMs (and on their implementation in the qgam package) which much alleviate this issue. In particular, we will show that the new methods lead to a significant decrease in computational time and to much lower memory requirements, but do not affect the accuracy of the fitted quantiles. While we will demonstrate the methods on regional solar production modelling, they are useful in a wide range of industrial and scientific applications.