After training heart.scale data set using RBF kernel, to find number of support vectors in each class I am writing following code. Am I doing any thing wrong ? as I am getting the same number of support vectors in each class.
supp_vect_index=model.sv_indices;
nt=size(supp_vect_index,1);
numTrain=size(train_labels);
for i=1:nt
for j=1:numTrain
if(supp_vect_index(i)==j)
supp_labels(i)=train_labels(j);
end
end
for i=1:nc
n(i)=0;
end
for i=1:nt
if(supp_labels(i)==1)
n(1)=n(1)+1;
elseif(supp_labels(i)==-1)
n(2)=n(2)+1;
end
end
Sorry! I forgot to mention in my question above that I am using LIBSVM with matlab interface, RBF kernel.
SVM always builds a linear separating hyperplane in feature space, hyperplane that is defined by the support vectors (and at the same distance of the two classes). If you set up the parameters in such a way that the number of support vectors is the minimum possible, and the two classes are balanced, the probability of having a similar number of support vectors is rather high.
Thanks Alberto.
Actually my data set is imbalanced even though I am getting same number of support vectors.I will check again by changing parameter values.
- Badges
- Science topic
- Similar topics
- Linguistics
- Composition Studies
- Writing