[Biojava-l] Re: Looking for a Fisher(like) Kernel

[Matthew Pocock]

On Thursday 01 December 2005 22:06, Todd Riley wrote:
> Hello,
>
> I have good news! I have fixed the bug in the BaumWelchTrainer class
> (hopefully the source in CVS will be updated soon).

Yay! What was it?

>
> Now that I am able to train my Profile HMM, I would like to feed my HMM
> into a Fisher Kernel to perform SVM training in order to find the proper
> scoring threshold for proper classification (ie - use SVM classification
> to set a barrier for my HMM log-odds scores).

Sounds like a plan...

>
> Has anyone implemented a Fisher Kernel (or one like it) for the BioJava
> SVM classes?  Any information here would be greatly appreciated.
>

I have not heard of one. However, I think you should be able to calcualte the 
needed numbers using code nearly identical to that in the BaumWelchTrainer. 
In fact, I have a sneeking suspicion that the ModelTrainer parameters after 1 
cycle of training (before updating the model!) are the raw numbers that the 
SVM fischer-kernel requires.

> Thanks,
> Todd

Matthew