Title: Selective inference for L2-boosting
Authors: David Ruegamer - LMU Munich (Germany) [presenting]
Sonja Greven - LMU Munich (Germany)
Abstract: The necessity for an explicit inference framework after model selection is due to the invalidity of classical inference after model selection. Following recent work on this topic, we address the issue of conducting valid inference when using boosting for variable selection and model fitting. We make use of a recently proposed inferential framework called selective inference, which corrects the inference after the selection procedure. After addressing several issues associated with an inference framework for component-wise boosting algorithms, we propose tests and confidence intervals for linear, grouped and penalized additive model estimates obtained after the L2-boosting selection process. By circumventing an explicit mathematical definition of the selective inference space, our algorithm is a flexible tool, which can also be used to incorporate resampling schemes such as cross-validation or stability selection into the inference framework.