Bagging, subagging and bragging for improving some prediction algorithms
Peter Bühlmann
February 2003
Abstract
Bagging (bootstrap aggregating), proposed by Breiman
[1], is a method to improve the predictive power of some special
estimators or algorithms such as regression or classification trees. First,
we review a recently developed theory explaining why bagging decision
trees, or also the subagging (subsample aggregating)
variant, yields smooth decisions, reducing the variance and mean squared
error. We then propose bragging (bootstrap robust
aggregating) as a new version of bagging which, in contrast
to bagging, is empirically demonstrated to improve also the MARS algorithm
which itself already yields continuous function estimates. Finally,
bagging is demonstrated as a "module" in conjunction with boosting for an
example about tumor classification using microarray gene expressions.
Download:
Compressed Postscript (99 Kb)
PDF (242 Kb).
Go back to the
Research Reports
from
Seminar für Statistik.