[Biojava-l] SVM and kernels

Matthew Pocock mrp@sanger.ac.uk
Fri, 04 Feb 2000 12:04:54 +0000


Dear Biojaverers,

(slightly off topic)

Support vector machines are the next big thing. They are easier to use &
generalise better than neural nets. Also, they are relatively simple to
implement. SVMs have been applyed to everything from hand-writing
recognition to gene functional class predicution from expression data.
We have the code already under org.biojava.stats.svm. Please pick it up
and play with it. This code is bleading edge development stuff.

I have found that the classifier seems to work if you set up all the
parameters carefuly. It is not the most intutative of interfaces at the
moment, but is workable. Take a look at
org.biojava.stats.svm.tools.Trainer and Classifier for an example of how
it works.

I am about to change the kernel objects so that they are easier to use.
In particular, the kernels that wrap up another kernel are going to all
inherit from an abstract NestedKernel. This should make it easier to
code up kernels, and also to build them in a java-bean aware enviroment.

If any of you have views about this code, I would love to hear about it.

Matthew