Does the Statistics Toolbox support hybrid kernels to train Support Vector Machines in MATLAB?

I see that the Statistics Toolbox R2014a provides a function called "svmtrain" to train an SVM. What kernels can I use to train it on my data? Can I use a hybrid kernel, i.e. a kernel made by mixing, for example, a linear kernel and a polynomial kernel?

Thanks!

Answer 1

[MathWorks Support Team]

The Statistics Toolbox R2014a offers the following kernels with the "svmtrain" function:

 - Linear

 - Quadratic

 - Polynomial

 - Gaussian Radial Basis Function

 - Multilayer Perceptron

 

Hybrid kernels are not readily available. However, there is an option to provide a custom kernel function using a function handle.

To train an SVM using a hybrid kernel, you would have to write the code for the hybrid kernel function. The kernel function must be of the form:

 

       function K = kfun(U,V)

 

where the returned value, "K", is a matrix of size M-by-N, and "U" and "V" have "M" and "N" rows respectively.

 

You could use the custom kernel "kfun" by specifying the "kernel_function" argument as follows:

 

       load fisheriris
       xdata = meas(51:end,3:4);
       group = species(51:end);
       svmStruct = svmtrain(xdata,group,'ShowPlot',true,'kernel_function',@kfun);

 

Here is an example of a hyperbolic tangent kernel that could be used with "svmtrain":

 

       function K = kfun(U,V)
       K = tanh(U*V');