When to use dual=False in LinearSVC? the document looks confusing

https://github.com/scikit-learn/scikit-learn/blob/fd237278e895b42abe8d8d09105cbb82dc2cbba7/sklearn/svm/_classes.py#L39

https://github.com/scikit-learn/scikit-learn/blob/fd237278e895b42abe8d8d09105cbb82dc2cbba7/sklearn/svm/_classes.py#L41

The default value for dual is True, which should correspond to the case where n_samples > n_features (usual case). But the document also says Prefer dual=False when n_samples > n_features., which looks really confusing to me.

So, which case we should use dual=False?

Thanks!

[whyisyoung]

In the primal formulation of linear SVC (i.e dual = False ), the optimisation variable is of dimension n_features. Whereas in the dual formulation (i.e dual = True ), the variable is of dimension n_samples. More importantly, the dual formulation requires the computation of an n_samplesxn_samples matrix . For this reason, when n_samples > n_features it is better to use dual = False .

See here for more details.

I think for most of the datasets, n_samples > n_features is the usual case. Then why not to set dual=False as the default setting?

[whyisyoung]

Thanks @gkevinyen5418. That makes a lot of sense to me. I will change to use dual=False in my case where n_samples > n_features and close this issue.