[Statlist] Talk by Torsten Hothorn

[Reinhard Furrer]

Wednesday, February 1, 2012, 16.15, Y27H25 (Irchel Campus)

Torsten Hothorn
Institut für Statistik, LMU München


Titel:
A fast and memory-efficient boosting implementation for generalized 
linear and additive model


Abstract:
Boosting can be seen as a very general functional approach to 
statistical model fitting. Its flexibility is extremely attractive also 
from a computational point of view, since a huge class of classical and 
modern statistical models can be fitted by such a procedure.
However, there is always a tradeoff between computational flexibility, 
generality and efficiency of a specific implementation. Focusing on 
generalized linear and additive models for problems where both the 
number of observations and exploratory variables may be in the millions, 
we present techniques to speed up computations and to reduce the memory 
footprint considerably.
Experiments suggest that high-dimensional linear models can be fitted by 
a componentwise boosting algorithm really fast (even faster than the 
lasso or elastic-net). Additive models for millions of observations can 
be fitted on standard desktop computers. Bootstrapping for model tuning 
and model inference benefits from these improvements as well.
The methodology is applicable for fitting a large class of classical and 
modern regression models, including robust regression, censored 
regression, quantile regression, or ordinal regression. The intrinsic 
variable and model selection procedure leads to parsimonous and 
interpretable models even in complex regression relationships.

-- 
______________________________________________________________________
University of Zurich                mailto:reinhard.furrer at math.uzh.ch
Institute of Mathematics                http://www.math.uzh.ch/furrer/
CH-8057 Zurich                       Tel/Fax: +41-(0)44-63-55843/55705