Date(s) : 23/03/2020 iCal
14 h 00 min - 17 h 00 min
Résumé :
A smooth, consistent regression tree and ensemble extensions
through RF and GBT
Joint work with S. Alkhoury (LIG), E. Devijver (LIG) and E. Gaussier (LIG)
Tree-based ensemble methods, as Random
Forests and Gradient Boosted Trees, have been
successfully used for regression problems in many
applications and research studies. We propose
here a generalization of regression trees, referred
to as smooth trees, that adapt to the smoothness
of the link function. By doing so, one consid-
ers that an observation, even though it belongs
to a particular region, can still be associated to
other regions with a certain weight that depends
on the distance between the observation and the
region. This generalization raises several difficult
questions, in particular regarding consistency. We
show here that smooth trees are indeed consistent,
a property that has not been established, as far
as we know, on previous proposals as soft trees.
We then show how smooth regression trees can
be used in different ensemble methods, namely
Random Forests and Gradient Boosted Trees. Ex-
periments conducted on several data sets further
illustrate the good behavior of smooth trees and
their ensemble extensions.
Emplacement
FRUMAM, St Charles (3ème étage)
Catégories