I am using libsvm in matlab interface. I want to find and store the value of discriminant function f(z)=sum(alphas*ys*k(xs ,z))+b. How can I do this?
Dear Krupal Parikh
The predict function of Libsvm gives three outputs by default:
>> [predictedLabelVec, overallAccuracy, probEstimates] = svmpredict(labelVec, featureVecs, svmModel, svmParams);
The third output, i.e. "probEstimates", gives a (somehow) normalized version of the values of discriminant function. I think you can use it as the discriminant values.
The above answer is correct. It is returning what you are seeking for.
Hi, you find here the program code under Matlab:
function [s,f] = svmdecide(samples,svmstruct)
% svmdecide: [s,f] = svmdecide(samples,svmstruct);
% To calculate the decision function for a classification with a support
% vector mahine trained and whoose parameters are stored in the svmstruct
% structure.
% the matrix samples contains the group of samples to be classified
% outputs are the decision function value vector f and its sign vector s.
%-------------------------------------------------------------------------%
% gets prameters from the structure:
sv = svmstruct.SupportVectors;
alpha = svmstruct.Alpha;
b = svmstruct.Bias;
kf = svmstruct.KernelFunction;
kfarg = svmstruct.KernelFunctionArgs;
scfact = svmstruct.ScaleData.scaleFactor;
scshift = svmstruct.ScaleData.shift;
% rescales samples according to the training:
for i = 1:size(samples,2);
samples(:,i) = scfact(i)*(samples(:,i)+scshift(i));
end
% calculates the decision:
f = -((feval(kf,sv,samples,kfarg{:})'*alpha(:))+b);
s = sign(f);
s(s==0) = -1;
%-------------------------------------------------------------------------%
Thanks to all.
Still wants to clarify more about my peroblem. My problem has 6 classes. After creating model file using RBF kernel , I want to use my own kernel in which I want to use value of decision function, which i am getting using standard RBF kernel. How do I get it from LIBSVM? I am using MATLAB interface of LIBSVM.
Dear Vijay Bhaskar,
In my problem, there are more than 70 features and 6 classes. It is multi classification problem. During training using one-to-one algorithm svm is doing training. So, it is difficult to extract dec-vaule.
- Badges
- Science topic