SVM classification with different kernels

I am using an SVM (SVM_train, Bioinformatics toolbox) to classify data, and I would like to have my final trained SVM models with different kernel functions. I didn't understand how to specify my own kernel maps: if I would like to use a Cauchy kernel defined as

k=(1/(1+(|u-v|^2/sigma))

where u,v are the vectors of the X data, and sigma is the parameter defined by the user (sigma=2.5). How I have to edit this matlab statement?

svmStruct = svmtrain(X,group,'Kernel_Function', 'rbf','RBF_Sigma', 2.5, 'Method', 'QP');

Thank you!

[Jeremy Huard]

Users using R2014b or newer should use the fitcsvm function for for one-class and binary classification or fitcecoc for multiclass classification instead.

There you can specify a custom kernel function by adding the 'KernelFunction','myfunction' name-value pair, where myfunction is the name of your function containing the kernel function definition.

Answer 1

[Ilya]

See this thread http://www.mathworks.com/matlabcentral/answers/57232-how-to-use-svmtrain-with-a-custom-kernel-in-matlab

Answer 2

[caterina]

Hi Ilya anche thank you for you answer. I read it that thread before writing and still I did not understand what it is wrong in my code. I created my kernel code that I saved in a folder 'pgm_qkda'. You can find the code below

clear all
clc;
addpath('/Users/Documents/Parallels/shared/D/KQDA/matlab2/pgm_qda');
x=[3 3 4 5 6 7 8 0.1 0.2
 3 4 3 5 7 7 5 0.5 0.6 
 7 7 2 3 3 4 5 0.4 0.5
 3 3 4 5 6 2 6 0.2 0.3];
group=[1,1, 2, 2]';
p1=2.5;
svmStruct = svmtrain(x,group,'Kernel_Function', @(u,v) kfun(u,v,p1), 'Method', 'QP');

The kfun function is defined as:

function kval = kfun(u,v,p1);
 dot=((u-v)*(u-v)')/p1;
 kval = 1/(1+dot);

Matlab gives me a warning message: Error using ==> svmtrain at 453 Error calculating the kernel function: Matrix dimensions must agree.

I do not understand where is my mistake. If you can help me I really appreciate.

[Ilya]

Replace

kval = 1/(1+dot);

with

kval = 1./(1+dot);

./ is for elementwise division. / is for matrix division.

Also, please do not post follow-up questions as answers to your own posts. Use comments for that.

[Fourth Sem Geethanjali Electrical and Electronics Engineering]

hi

svmStruct = svmtrain(x,group,'Kernel_Function', @(u,v) kfun(u,v,p1), 'Method', 'QP');

how will the function call change if i use fitcsvm instead of svmtrain.

can you please help me.

Answer 3

[Sandy]

 Hi Caterina,

Just want to check whether you are able to rectify tha above issues. I am also getting the same error. Will you post me the code for above custom kernel.

Thanks in advance.

Answer 4

[muqdad aljuboori]

Hi I very interested in this discussion I tried to apply the new funcion in my project but it doesnt work that what i did

 SVR1 = fitrsvm(TrainInputs,TrainTargets,...
    'KernelFunction', @(u,v) kfun3(u,v) ,...
    'KernelScale','auto','Standardize',true);__

 and the error is

Error using classreg.learning.modelparams.SVMParams.make (line 225) You must pass 'KernelFunction' as a character vector.

Error in classreg.learning.FitTemplate/fillIfNeeded (line 598) this.MakeModelParams(this.Type,this.MakeModelInputArgs{:});

Error in classreg.learning.FitTemplate.make (line 124) temp = fillIfNeeded(temp,type);

any suggestion ??? thanks

[Jeremy Huard]

You should specify the kernel function as char array:

 SVR1 = fitrsvm(TrainInputs,TrainTargets,...
    'KernelFunction', 'kfun3',...
    'KernelScale','auto','Standardize',true);