quadprog function for non-separable data-set

I want to implement svm for two sets of non-separable cases using svm primal form(without using built-in functions). In the minimize quadratic function, I have 1/2 norm(w)^2 + C*summation over all feature vectors(epsilon). If I use quaprog function, I am not sure if I need to have the vector 'f' since I have only one quadratic term in w. But without using 'f' the minimizing expression is not complete.

Could someone please help on how to build the matrices of 'quadprog' function for non-separable svm primal case?

Thanks so much

Answer 1

[Matt J]

Just set f=zeros(size(w)), if there is no linear term in your objective function.

[rini]

Thak you so much for the answer, but if I set 'f' to be zero doesn't the problem becomes similar to the linearly separable case? For non-separable cases do we totally neglect this factor?

Thanks.

[Matt J]

I have no background in SVM, but based on wiki:SVM Primal Form, you indeed have f=0 and you also have non-trivial linear constraints that you have not mentioned. The constraints render the problem nonseparable, as long as they're not simple bound constraints.

[Matt J]

If your constraints are simple bound constraints, the problem will be separable purely by virtue of the quadratic term being norm(w)^2/2 regardless of whether f=0 or not. The linear term is always separable.

[rini]

thank you so much

[Matt J]

Accept-clicking my answer is all the thanks I need ;)