Date:         Sat, 3 Feb 2007 18:17:42 GMT
Reply-To:     Phil Sherrod <phil.sherrod@REMOVETHISSANDH.COM>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Phil Sherrod <phil.sherrod@REMOVETHISSANDH.COM>
Subject:      Re: RuleFit versus MART or TreeNet
Comments: To: sas-l@uga.edu

On 30-Jan-2007, dlouiszhao@GMAIL.COM (Daqing Zhao) wrote:

> I wonder if anyone has used RuleFit and how it compares to TreeNet in terms
> of accuracy, robustness with respective to outliers and missing data, etc.

I don't have any experience with RuleFit, but I have a lot of positive
experience with TreeBoost [TreeNet(tm)].  I've benchmarked TreeBoost against
many types of models including neural networks, SVM and decision tree forests.
TreeBoost usually comes out first or second, and it rarely performs poorly.  It
is much more consistent than SVM, and the training times with large datasets
are much better.  I've got customers using DTREG to build TreeBoost models with
500+ predictor variables and several hundred thousand training rows.

Here is a page with benchmarks comparing TreeBoost with other methods:
http://www.dtreg.com/benchmarks.htm

--
Phil Sherrod
(phil.sherrod 'at' sandh.com)
http://www.dtreg.com  (Decision trees, Neural networks and SVM modeling)
http://www.nlreg.com  (Nonlinear Regression)