You can also put out lots of clusters and use cluster membership as the
features for a classifier.

There was a discussion here (or possibly on the dev@mahout list) on this
topic several weeks ago.  Search the archives for "iris" and my name.



On Fri, Jan 24, 2014 at 8:46 AM, Angus Macnab <[email protected]>wrote:

> You can do supervised learning by outputing the clusters and labeling them
> 0-9.
>
> > On Jan 23, 2014, at 10:34 PM, Tharindu Rusira <[email protected]>
> wrote:
> >
> > On Fri, Jan 24, 2014 at 9:50 AM, Angus Macnab <[email protected]
> >wrote:
> >
> >> This is a pretty classic machine learning problem and can be handled
> with
> >> several different algorithms.  Logistic regression is the obvious
> choice,
> >> but clustering algorithms will work fine also.  Just decompose the
> pixels
> >> into a really long vector and train your algorithm with the input-output
> >> pairs.  You can get 100% accuracy on this pretty easily if you are
> careful
> >> with your bias-variance decomposition.  This is a fun one for neural
> >> networks too!
> >>
> >> Essentially any machine learning book will delve into greater detail on
> >> this as the US postal digit data has been around for a long time.  I
> think
> >> Kaggle even had this as a training exercise for a while, so there's
> >> probably a ton of discussion of various methods and algorithms on their
> >> message boards.
> >>
> >> For kicks why don't you compare k-means clustering to logistic
> regression
> >> using Mahout?
> > Hi Angus, Chameera's requirement is to classify handwritten digits, so
> > could you please explain how could K-means clustering be helpful in this
> > scenario? Of course it would find different clusters but this is still a
> > classification problem. Please correct me if I'm wrong.
> >
> > Thanks,
> >
> >
> >>
> >> -Angus
> >>
> >>
> >>
> >>
> >> On Thu, Jan 23, 2014 at 8:00 PM, Chameera Wijebandara <
> >> [email protected]> wrote:
> >>
> >>> Hi,
> >>>
> >>> I am trying to classify handwritten digits using mahout classification.
> >> Any
> >>> suggestion to come up with good solution?
> >>>
> >>> --
> >>> Thanks,
> >>>    Chameera
> >
> >
> >
> > --
> > M.P. Tharindu Rusira Kumara
> >
> > Department of Computer Science and Engineering,
> > University of Moratuwa,
> > Sri Lanka.
> > +94757033733
> > www.tharindu-rusira.blogspot.com
>

Reply via email to