[Biojava-l] SVM regression

[Matthew Pocock]

Hi Mark,

You can convert HMMs into kernels (assuming that they are fully 
symetrical pairwise aligners) using the biojava DP package. You will 
have to write a bit of code to capture the agregates of the 
forwards/backwards scores into a ModelTrainerContext (see 
BaumWelchTrainer for an example of how to do this). From these numbers 
you should be able to calculate the apropreate matricies. The colt is a 
good Java matrix library. I have found that using raw HMM forwards 
scores (either probabilities, log probabilities, or either of the above 
projected onto a unit sphere) gives realy crappy results, as most 
sequences don't resemble each other very much. Also, pay some attention 
to the architecture of the HMM used to align the sequences. This acts as 
in the flavour of a prior over which types of sub-sequences are 
informative basies in the suffix-tree space implicitly constructued by 
the kernel. If this space is the 'wrong' shape, then vast numbers of 
support vectors will be required to define a suitable plane.

All the best,

Matthew

Mark Schreiber wrote:

> Hi -
> 
> I am thinking of using SVM's and I have a question about the SVM package.
> Is there a way of converting an HMM into a kernel as per the technique of
> Jaukkola and Haussler?
> 
> Mark
> 
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> Mark Schreiber			Ph: 64 3 4797875
> Rm 218				email mark_s@sanger.otago.ac.nz
> Department of Biochemistry	email m.schreiber@clear.net.nz
> University of Otago		
> PO Box 56
> Dunedin
> New Zealand
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> 
> _______________________________________________
> Biojava-l mailing list  -  Biojava-l@biojava.org
> http://biojava.org/mailman/listinfo/biojava-l
> 
> 
>