Abstract
BACKGROUND: Statistical boosting is a computational approach to select and estimate interpretable prediction models for high-dimensional biomedical data, leading to implicit regularization and variable selection when combined with early stopping. Traditionally, the set of base-learners is fixed for all iterations and consists of simple regression learners including only one predictor variable at a time. Furthermore, the number of iterations is typically tuned by optimizing the predictive performance, leading to models which often include unnecessarily large numbers of noise variables. RESULTS: We propose three consecutive extensions of classical component-wise gradient boosting. In the first extension, called Subspace Boosting (SubBoost), base-learners can consist of several variables, allowing for multivariable updates in a single iteration. To compensate for the larger flexibility, the ultimate selection of base-learners is based on information criteria leading to an automatic stopping of the algorithm. As the second extension, Random Subspace Boosting (RSubBoost) additionally includes a random preselection of base-learners in each iteration, enabling the scalability to high-dimensional data. In a third extension, called Adaptive Subspace Boosting (AdaSubBoost), an adaptive random preselection of base-learners is considered, focusing on base-learners which have proven to be predictive in previous iterations. Simulation results show that the multivariable updates in the three subspace algorithms are particularly beneficial in cases of high correlations among signal covariates. In several biomedical applications the proposed algorithms tend to yield sparser models than classical statistical boosting, while showing a very competitive predictive performance also compared to penalized regression approaches like the (relaxed) lasso and the elastic net. CONCLUSIONS: The proposed randomized boosting approaches with multivariable base-learners are promising extensions of statistical boosting, particularly suited for highly-correlated and sparse high-dimensional settings. The incorporated selection of base-learners via information criteria induces automatic stopping of the algorithms, promoting sparser and more interpretable prediction models.