Hello,
Thanks for your reply. I do agree with the change of C values the
solution would be unique and optimal..but I guess what I didn't specify
clearly enough was this-
When one applies the SVM..one also changes other parameters such as
lambda( or the conditioning parameter for quadratic programming) and
epsilon(the width of the e-tube in the case of svm regression) and the
different values necessary for a particular kernel to finally get the best
model. So does one get the optimal and unique solution with each
different value or is there a right value for each of these parameters
after which one can be guaranteed the unique and optimal solution?
Regarding the microrray data- My opinion also matches yours. There are
many papers out there that do gene expression analysis and though they
always
mention the fact that the number of samples is less than the number of
features, they hardly ever mention how many samples they had. Moreover,
when you have many classes and not a binary classification scenario, it
gets very expensive to just have 2-3 experiments per class (not to
mention the failed experiments), so I can
understand where the limitation is. Has anyone had experience
in dealing with this kind of a problem?
Thank You.
Sincerely,
Monika Ray
***********************************************************************
The sweetest songs are those that tell of our saddest thought...
Computer Science, Washington University, St. Louis, MO, USA
***********************************************************************
On Thu, 9 Sep 2004, Junshui Ma wrote:
> > SVM has been known to have a unique and optimal solution...does this mean
> > with each new C (and other parameter values ) value you get a unique and
> > optimal solution? or is it
> > for the right C/parameters value you will get the unique and optimal
> > solution?
>
> SVM has global optimal solution(s) because its training procedure is
> essentially a quadratic problem (with linear constraints).
> This nature does not change for any given C values.
>
> In some cases ( due to the special form of kernel matrix), SVM may not have
> a "unique" optimal solution.
>
> > Also, its well known that microarray data have high number of features but
> > not many samples. I have data where each class was repeated 3 times.
> > That is -
> >
> > For Class 1 - 3 independent experiments (out of which 1 expt. was too
> > noisy and had to be thrown out..so you end up with just 2 experimenst of
> > valid results)
> >
> > For Class 2 - 3 independent experiments (out of which 1 expt. was too
> > noisy and had to be thrown out..so you end up with just 2 experimenst of
> > valid results)
> >
> > For class 3 - 3 independent experiments (all valid)
> >
> > and so on for 2 more classes - therefore have 5 classes all in all.
> >
> > Now My question is -
> >
> > I don't think there are enough examples per class. So is it wise to train
> > an SVM on this?
>
> Although you still can apply SVM to this dataset and get something out of
> it,
> I am not sure your result will be reliable and/or reasonable.
> In fact, I don't know if it is possible for you to get *any* reliable
> results out of this kind of dataset.
> I really want to learn more about others' opinions in this kind situation.
>
> jm
>
> -------------------------------------------
> Junshui Ma, Ph. D.
> 614-688-4893 (office)
> Room 350A, Ohio Supercomputer Center
> 1224 Kinnear Rd., Columbus, OH 43212
> http://www.osc.edu/~junshui
>
|