What are the differences between SVC, NuSVC, and LinearSVC?
Please shed some light.
$\begingroup$
$\endgroup$
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
3
In scikit SVC and nuSVC are mathematically equivalent with both methods based on the library libsvm. The main difference is that SVC uses the parameter C while nuSVC uses the parameter nu.
LinearSVC is based on the library liblinear. As the documentation says, LinearSVC is similar to SVC with parameter kernel='linear', but liblinear offers more penalties and loss functions in order to scale better with large numbers of samples. Please check out this question and this question for more details.
-
1$\begingroup$ But why does it give different results for
NuSVCand ```SVC`` on same data`? $\endgroup$– DeshwalDec 18, 2019 at 12:07 -
$\begingroup$ @Deshwal It's difficult to say without seeing the code. My guess would be that the parameters used in each case are not equivalent. Apparently the reparametrization is "not that easy to calculate" according to this answer that references this paper. $\endgroup$ Dec 18, 2019 at 18:02
-
$\begingroup$ It is explained in the user guide. NuSVC's
vparameter serves as both "to upper bound the fraction of margin errors and a lower bound of the fraction of support vectors." I interpret it as e.g. at most 2% margin error allowed, and at least 2% examples to use as support vectors. $\endgroup$– ShernMay 28, 2020 at 9:19