Date: Sat, 21 May 2011 19:23:41 -0400
Reply-To: oloolo <dynamicpanel@YAHOO.COM>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: oloolo <dynamicpanel@YAHOO.COM>
Subject: Re: How many of you want a SAS implementation of ElasticNet
The math is not that difficult, actually and it is absolutely doable, but of
course in the way SAS program should be written, not directly follow the
decending coordinate method shoown in their JSS paper.
I've been trying GLMNET package in R in a couple of projects, and the
results were superior to other methodologies, including Ridged LDA, GDA and
SVM. It demonstrated consistently high AUC while low MSE(AUC) in CV.
The advantage of Elastic Net over LASSO is two folds:
1. Dealing with P>>N cases and select more than N variables
2. Force penalized non-corner solution when multicollinearity happens, which
will improve predictive power
The advantage of Elastic Net over ridge is that in some cases, you can
achieve sparseness on covariance like in LASSO (which in my experience
rarely happened, though)
But you do have 2 more tuning parameter, making it more flexible than above two.
On Sat, 21 May 2011 14:19:39 -0400, Wensui Liu <liuwensui@GMAIL.COM> wrote:
>if what you referred is one by Zou and Hastie, I definitely see its
>value primarily in the use case of variable selection in the industry.
>i just also curious to know if it is worth the effort to get that deep
>and show the extra business benefit in model development.
>On Thu, May 19, 2011 at 1:34 PM, oloolo <email@example.com> wrote:
>> ElasticNet is one of the hottest data mining algorithms/ variable selection
>> methodologies in recent years.
>> Just out of curiousity, how many of SASLORs want an implementation in SAS?