Thank you very much Ted.  This clears up a lot of the terminology for me as
the concept of ridge was new to me as of this week.
Question: When you state:

b) the --categories is there for the command line version which is mostly
for demo purposes

... I am unclear as to what you mean.  I thought all of Mahout ran from the
command line. What am I missing out on?

Keith

On Thu, Sep 29, 2011 at 3:46 AM, Ted Dunning <[email protected]> wrote:

> a) multinomial is supported
>
> b) the --categories is there for the command line version which is mostly
> for demo purposes
>
> c) ridge regression originally meant L2 regularization.  Nowadays people
> use
> it to describe either L1 or L2, but Lasso would be a better word for L1.
>
> d) regularization is independent of link function.  Mahout only supports
> inverse logistic link function (i.e. logistic regression).  Other links are
> easy to install.  Linear regression is very easy to support, but since
> there
> is a strong initial solution (using Cholesky decomposition = L2
> regularization), I would recommend a from-scratch implementation.
>
> e) Hector's work has added some additional flexibility, mostly untried as
> yet.
>
> On Thu, Sep 29, 2011 at 12:01 PM, Dmitriy Lyubimov <[email protected]
> >wrote:
>
> > If by ridge estimator you imply l1 regularization, afaik you can choose
> > from
> > l1 and l2 gularization.
> >
> > Also I guess ridge implies identity link function, but I don't think
> mahout
> > is limited just to ridge and lasso.
> >
> > Ted knows for sure.
> > On Sep 28, 2011 7:27 PM, "Keith Thompson" <[email protected]>
> wrote:
> > > Thanks Dmitriy. I also found the Mahout in Action book and checked
> there.
> > > In one of the examples it seems there is a --categories option which
> > allows
> > > the user to specify the number of categories for classification
> purposes
> > > (although all of the examples in the text are binary classifications).
> I
> > am
> > > curious though as to whether the algorithm employed in Mahout uses
> ridge
> > > estimators or not.
> > >
> > > Keith
> > >
> > > On Wed, Sep 28, 2011 at 10:18 PM, Dmitriy Lyubimov <[email protected]
> > >wrote:
> > >
> > >> As far as I know, Mahout implementation does multinomial
> classification.
> > >> On Sep 28, 2011 8:47 AM, "Keith Thompson" <[email protected]>
> > wrote:
> > >> > Is the version of the logistic regression classifier in Mahout for
> > >> binomial
> > >> > classification only or can it be used for multinomial as well?
> > >> >
> > >> > Regards,
> > >> > Keith
> > >>
> > >
> > >
> > >
> > > --
> > > *Keith Thompson*
> > > Graduate Research Associate, Xerox Corporation
> > > SUNY Research Foundation
> > > Dept. of Systems Science and Industrial Engineering
> > > Binghamton University
> > > work: 585-422-6587
> > > cell: 607-793-6203
> >
>



-- 
*Keith Thompson*
Graduate Research Associate, Xerox Corporation
SUNY Research Foundation
Dept. of Systems Science and Industrial Engineering
Binghamton University
work: 585-422-6587
cell: 607-793-6203

Reply via email to