[R-sig-hpc] svm for Bioinformatics

[Luis Cebamanos]

Hi Bernd,

Thank you very much for your help and interest.

> could you state a bit more specifically what your questiona regarding 
> SVM performance are?
>

As you have correctly mentioned, the dataset that I have belongs to 
bioinformatics where the number of genes (rows) could be larger than 
300000 and the number of observations (in this case patients, columns ) 
will never be larger than 1000. I have duplicated this matrix and it is 
still very fast to calculate the svm in R. My goal is to speed up the 
svm but unfortunately I can't find the proper dataset that allows me to 
profile it and see the areas where the svm spends more time.
Originally, this idea came from bioinformaticians, however it might be 
applied to any other areas. I have also seen the svm may have quite a 
lot of different parameters, and I was wandering if some of them, which 
are really useful would slow down the application and would be worth it 
to study its parallelization.
Finalising, I would like to know if svm in R is not fast enough, and in 
this case, under which situation (parameters, dataset, etc).

Does this explain a bit better my intentions?

Best,
Luis


> Also, because you mention bioinformatics apüplications, I assume you 
> work in a setting where p >> n, and n is often quite small, is that 
> essentially correct?
>
> Bernd
>
> On 13.06.2012 12:45, Luis Cebamanos wrote:
> > Hi all,
> >
> > I am studying the svm function (included in e1071 package) for 
> bioinformatics in
> > order to improve its performance. From an R users survey, it was 
> mention as one
> > of the functions to improve the performance in bioinformatics.
> > After profiling the application, we have not found difficulties 
> running large
> > datasets and I was wandering if someone often uses svm and could 
> give us some
> > information about its performance.
> >
> > Best regards,
> > Luis
> >
>
>

-- 
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.