svm

This is Anton Schwaighofer's SVM toolbox for MATLAB. It used to be hosted by Anton on line but the page is down so we've added it here.

Support Vector Machine toolbox for Matlab Version 2.51, January 2002

Contents.m contains a brief description of all parts of this toolbox.

Main features are:

Except for the QP solver, all parts are written in plain Matlab. This guarantees for easy modification. Special kinds of kernels that require much computation (such as the Fisher kernel, which is based on a model of the data) can easily be incorporated.
Extension to multi-class problems via error correcting output codes is included.
Unless many other SVM toolboxes, this one can handle SVMs with 1norm or 2norm of the slack variables.
For both cases, a decomposition algorithm is implemented for the training routine, together with efficient working set selection strategies. The training algorithm uses many of the ideas proposed by Thorsten Joachims for his SVMlight. It thus should exhibit a scaling behaviour that is comparable to SVMlight.

This toolbox optionally makes use of a Matlab wrapper for an interior point code in LOQO style (Matlab wrapper by Steve Gunn, LOQO code by Alex Smola). To compile the wrapper, run mex loqo.c pr_loqo.c Make sure you have turned on the compiler optimizations in mexopts.sh The LOQO code can be retrieved from http://www.kernel-machines.org/code/prloqo.tar.gz The wrapper comes directly from Steve Gunn.

This program is released unter the GNU General Public License. See License.txt for details.

Changes in version 2.51:

fixed bug in SVMTRAIN that prevented correct initialisation with NET.recompute==Inf

Changes in version 2.5:

Handling of multi-class problems with ECOC
NET.recompute is set to Inf by default, thus all training is done incrementally by default.
Handling the case of all training examples being -1 or +1 correctly

Changes in version 2.4:

Better selection of the initial working set
Added workaround for a (rare) Matlab quadprog bug with badly conditioned matrices
There is now a new kernel function 'rbffull' where a full matrix ("covariance matrix") C may be put into an RBF kernel: K(X1,X2) = exp(-(X1-X2)'C(X1-X2))

Changes in version 2.3:

slightly more compact debug output

Changes in version 2.2:

New default values for parameter qpsize that make the whole toolbox much faster
Workaround for a Matlab bug with sparse matrices
Changed the definition of the RBF-Kernel: from |x-y|^2/(2ninparam^2) to |x-y|^2/(nin*param). This means that all parameter settings for old versions need to be updated!
A few minor things I can't remember

Changes in version 2.1: Fixed a nasty bug at the KKT check

Changes in version 2.0: All relevant routines have been updated to allow the use of a SVM with 2norm of the slack variables (NET.use2norm==1).