|
Abstract:
|
Classically, boosting or functional gradient descent algorithms for optimizing various empirical risk functions have been implemented using relatively complex base-learners. Recently, boosting algorithms for fitting generalized linear or additive models have been suggested. The key innovation is the application of componentwise linear models or smoothing splines which allows us for a reformulation in terms of classical linear or additive models. In the former case, the regression coefficients can be interpreted in the usual way. Moreover, those boosting algorithms have been demonstrated to be useful for variable selection in high-dimensional situations. We briefly sketch the underlying theory and demonstrate how to actually fit regression models with many covariates in the R system for statistical computing using the `mboost' add-on package.
|
