MATLAB/MEX Interface to SVM^light

[download] - download links and instructions

Updated July 6, 2005

• Accepts MATLAB's "sparse" matrices as input.
• If sparse matrice given as input, support vectors returned as sparse output.
• Conditional compilation of memory allocation tracker to prevent double-free() on a block ( use -DMEX_MEMORY and see mem_clean.h for hash map bucket size). This feature is off for UNIX systems, and on for Windows systems.
• Several bug fixes requested by users (e.g. missing status output, missing variables in the model).
• Improved(?) installation procedure.
• Support for SVM-Light's regression (-z r) and ranking (-z p).

MEX-SVM is an interface between MATLAB and SVM^light , a powerful and efficient SVM package developed by Thorsten Joachims. The primary goal of this package is to have a simple and efficient interface between the two packages.

Why Another MEX-SVM Package?

• SVM_^light is extremely powerful, fast, and stable
• SVM_^light code organized well and lends itself to extension
• SVM_^light 6.0 includes new features, and needed a good MATLAB front-end
• MATLAB is a natural environment for data exploration
• MEX-SVM interface of MATLAB allows in-memory direct copies of data, eliminating need to create temporary files and creating external shells to execute code

How to Use the Package?

The interface is extremely simple. There are detailed installation and usage instructions in the distributions. The following is a summary of the capabilities of this package.

There are two MATLAB functions, svmlearn and svmclassify. The first function, svmlearn, takes three arguments. The first two arguments, x and y,are matrices corresponding to the training data (x) and labels (y). The third parameter is a string composed of the parameters of the command-line version of the SVM^light software. The output of svmlearn is a MATLAB structure derived from the fields of the SVM^light model. See [model] for a description of each parameter and its corresponding SVM^light source parameter

The second MATLAB function, svmclassify, also requires three input parameters. The first two are the x and y parameters of data to be classified. The third parameter is the model structure from a previous call to svmlearn. Two outputs are produced, the error rate and an array of predictions.

Following is a sample session demonstrating the use of the package. In the following example, assume that X is a 400x2 MATLAB matrix corresponding to the training data, and Y is a 400x1 MATLAB matrix corresponding to the class labels. The given parameter list sets SVM^light to use the RBF kernel (-t 2), set gamma to 0.3 (-g 0.3), and set C to be 0.5 (-c 0.5). Any of the standard SVM^light command line arguments may be specified, see SVM^light documentation for details. (note: one useful parameter is '-v 0' to turn SVM^light output off completely.)

>> model = svmlearn(x,y,'-t 2 -g 0.3 -c 0.5')
Optimizing...................................................................
.............................................................................
.............................................................................
.............................................................................
.....
 Checking optimality of inactive variables...done.
 Number of inactive variables = 335
done. (304 iterations)
Optimization finished (46 misclassified, maxdiff=0.00073).
Runtime in cpu-seconds: 0.13
Number of SV: 272 (including 262 at upper bound)
Computing XiAlpha-estimates...done
Runtime for XiAlpha-estimates in cpu-seconds: 0.00
Number of kernel evaluations: 79422

model = 

           sv_num: 273
      upper_bound: 262
                b: -0.1069
         totwords: 2
           totdoc: 400
        loo_error: -1
       loo_recall: -1
    loo_precision: []
         xa_error: 60.2500
        xa_recall: 44.7005
     xa_precision: 44.4954
          maxdiff: 7.3188e-004
       r_delta_sq: 1.8625
      r_delta_avg: 0.8535
     model_length: 7.6915
             loss: 148.5834
            vcdim: 107.3005
            alpha: [400x1 double]
      lin_weights: []
            index: [400x1 double]
           supvec: [273x2 double]
      kernel_parm: [1x1 struct]


>> [err , predictions ] = svmclassify(x,y,model);
>> err

err =

    0.1150

>> 

© Tom Briggs,
All Rights Reserved.

This page last modified: Wednesday, July 06, 2005