Search

Component-wise gradient boosting algorithms are popular for their intrinsic variable selection and implicit regularization, which can be especially beneficial for very flexible model classes. When estimating generalized additive models for location, scale and shape (GAMLSS) by means of a component-wise gradient boosting algorithm, an important part of the estimation procedure is to determine the relative complexity of the submodels corresponding to the different distribution parameters. Existing methods either suffer from a computationally expensive tuning procedure or can be biased by structural differences in the negative gradients' sizes, which, if encountered, lead to imbalances between the different submodels. Shrunk optimal step lengths have been suggested to replace the typical small fixed step lengths for a non-cyclical boosting algorithm limited to a Gaussian response variable in order to address this issue. In this article, we propose a new adaptive step length approach that accounts for the relative size of the fitted base-learners to ensure a natural balance between the different submodels. The new balanced boosting approach thus represents a computationally efficient and easily generalizable alternative to shrunk optimal step lengths. We implemented the balanced non-cyclical boosting algorithm for a Gaussian, a negative binomial as well as a Weibull distributed response variable and demonstrate the competitive performance of the new adaptive step length approach by means of a simulation study, in the analysis of count data modeling the number of doctor's visits as well as for survival data in an oncological trial., Comment: 34 pages, 26 figures

Boosting methods are widely used in statistical learning to deal with high-dimensional data due to their variable selection feature. However, those methods lack straightforward ways to construct estimators for the precision of the parameters such as variance or confidence interval, which can be achieved by conventional statistical methods like Bayesian inference. In this paper, we propose a new inference method "BayesBoost" that combines boosting and Bayesian for linear mixed models to make the uncertainty estimation for the random effects possible on the one hand. On the other hand, the new method overcomes the shortcomings of Bayesian inference in giving precise and unambiguous guidelines for the selection of covariates by benefiting from boosting techniques. The implementation of Bayesian inference leads to the randomness of model selection criteria like the conditional AIC (cAIC), so we also propose a cAIC-based model selection criteria that focus on the stabilized regions instead of the global minimum. The effectiveness of the new approach can be observed via simulation and in a data example from the field of neurophysiology focussing on the mechanisms in the brain while listening to unpleasant sounds.

Tuning of model-based boosting algorithms relies mainly on the number of iterations, while the step-length is fixed at a predefined value. For complex models with several predictors such as Generalized Additive Models for Location, Scale and Shape (GAMLSS), imbalanced updates of predictors, where some distribution parameters are updated more frequently than others, can be a problem that prevents some submodels to be appropriately fitted within a limited number of boosting iterations. We propose an approach using adaptive step-length (ASL) determination within a non-cyclical boosting algorithm for GAMLSS to prevent such imbalance. Moreover, for the important special case of the Gaussian distribution, we discuss properties of the ASL and derive a semi-analytical form of the ASL that avoids manual selection of the search interval and numerical optimization to find the optimal step-length, and consequently improves computational efficiency. We show competitive behavior of the proposed approaches compared to penalized maximum likelihood and boosting with a fixed step-length for GAMLSS models in two simulations and two applications, in particular for cases of large variance and/or more variables than observations. In addition, the idea of the ASL is also applicable to other models with more than one predictor like zero-inflated count model, and brings up insights into the choice of the reasonable defaults for the step-length in simpler special case of (Gaussian) additive models.