B1016
Title: Gradient boosting for GAMLSS using adaptive step lengths
Authors: Alexandra Daub - Georg-August-Universität Göttingen (Germany) [presenting]
Andreas Mayr - University of Bonn (Germany)
Boyao Zhang - University of Goettingen (Germany)
Elisabeth Bergherr - Georg-August-Univerität Göttingen (Germany)
Abstract: In order to benefit from the known advantages of machine learning methods, component-wise gradient boosting algorithms are used for estimating statistical models, i.e. generalized additive models for location, scale and shape (GAMLSS). Estimating GAMLSS by means of a non-cyclical gradient-based boosting algorithm with fixed step lengths can however result in imbalanced submodel updates and long run times. Optimal step lengths have been shown to solve these issues. A new way for obtaining adaptive step lengths is proposed based on algorithm intrinsic information and a non-cyclical boosting algorithm for GAMLSS is implemented with the different step length options for normal, negative binomial and Weibull distributed response variables. A simulation study as well as the application on real-world data sets show that the new adaptive step length yields similar results as a numerically obtained optimal step length while reducing the run time considerably.
