I can perhaps add this too the scale_c example as well
2012/8/21 Alexis Mignon
> It is written in the documentation of the loss parameter:
>
> string, 'l1' or 'l2' (default='l2')
> Specifies the loss function. 'l1' is the hinge loss (standard SVM)
> while 'l2' is the squared hinge loss.
It is written in the documentation of the loss parameter:
string, 'l1' or 'l2' (default='l2')
Specifies the loss function. 'l1' is the hinge loss (standard SVM)
while 'l2' is the squared hinge loss.
but clearly it can be passed unseen with a quick reading.
2012/8/21 Andreas Mueller
>
Hi Alexis.
That must be it! Crazy.
We should definitely point this out in the docs somewhere.
Not sure how I overlooked that.
Thanks a lot!
Andy
On 08/21/2012 10:52 AM, Alexis Mignon wrote:
Another big difference is that the default loss function in Linear SVC
is quadractic with distance to the
Another big difference is that the default loss function in Linear SVC is
quadractic with distance to the margin (loss='l2') while it is linear in
SVC.
2012/8/12 Mathieu Blondel
>
>
> On Sun, Aug 12, 2012 at 6:53 PM, Andreas Mueller > wrote:
>
>> Does any one have an explanation for that?
>> Bt
On Sun, Aug 12, 2012 at 6:53 PM, Andreas Mueller
wrote:
> Does any one have an explanation for that?
> Btw, I am using the sparse versions to do some text classification.
>
One difference is that SVC fits the intercept directly (without using the
dummy feature trick). So the intercept is not pena
Hi everybody.
Yesterday I noticed big differences in performance between SVC with
linear kernel and LinearSVC.
I vaguely remember there was an issue about that, but can't find it any
more.
I tried to set the stopping criterion very strict but still I saw a big
difference.
Does any one have an e
