B0533
Title: Stability selection for boosted generalized additive models for location scale and shape
Authors: Janek Thomas - Ludwig-Maximilians-University Munich (Germany) [presenting]
Andreas Mayr - University of Bonn (Germany)
Matthias Schmid - University of Erlangen-Nuremberg (Germany)
Bernd Bischl - LMU Munich (Germany)
Benjamin Hofner - Paul-Ehrlich-Institut (Germany)
Abstract: A new algorithm is proposed to incorporate stability selection for boosting generalized additive models for location, scale and shape (GAMLSS). In one application, a negative binomial hurdle model was fitted to handle excess zeros, overdispersion, non-linearity and spatiotemporal structures to investigate the abundance of seabirds in Massachusetts. In a second step, stability selection, an increasingly popular way to obtain stable sets of covariates while controlling the false discovery rate (FDR), was applied. The model is fitted repeatedly to subsampled data and variables with high selection frequencies are extracted. This lead to a fundamental problem with boosted GAMLSS, where in every boosting iteration, the algorithm sequentially selects the best fitting effect for each distribution parameter. Thus, it is currently not possible to stop fitting individual parameters as soon as they are sufficiently modeled. To solve this problem, we developed a new approach to fit boosted GAMLSS. Instead of updating all distribution parameters in each iteration, we only update the parameter which leads to the biggest reduction in loss. With this modification, stability selection can be applied. Furthermore, optimizing the tuning parameters of boosting is reduced from a multi- to a one-dimensional problem. The performance of the algorithm is evaluated in a simulation study and the application is demonstrated for the seabirds data, selecting stable predictors while controlling the FDR.
Journal