TR on fast computation of Leave-One-Out error in SVM algorithms

[S. Sathiya Keerthi]

A New Technical Report... 

Two Efficient Methods for Computing Leave-One-Out Error in SVM Algorithms 

S. Sathiya Keerthi, Chong Jin Ong and Martin M.S. Lee 
National University of Singapore


Abstract:

We propose two new methods for the efficient computation of
the leave-one-out (LOO) error for SVMs. The first method is based
on the idea of exact penalty functions while the second method
uses a new initial choice for the alpha variables. These methods
can also be extended to the recomputation of SVM solutions when more
than one example is omitted and the computation of LOO error for 
nu-SVM.  Recently, Lee and Lin pointed out that a loose stopping 
criteria can be exploited to speed up LOO computations for SVM.  
This fact, combined with our proposed methods allows for the efficient
computation of the LOO error.

To download a gzipped postscript file containing the report, go to:
http://guppy.mpe.nus.edu.sg/~mpessk/svm.shtml