On 01/08/2012 11:29 PM, Olivier Grisel wrote:
> 2012/1/8 Andreas<[email protected]>:
>    
>> Hey everybody.
>> @larsmans (my personal hero for the day) started refactoring the SVM
>> class structure here:
>> https://github.com/larsmans/scikit-learn/commits/refactor-svm
>> after some discussion here:
>> https://github.com/scikit-learn/scikit-learn/issues/253
>> and somewhat related here:
>> https://github.com/scikit-learn/scikit-learn/issues/100
>>
>> The bottom line is: the SVM class structure is not as nice as one might
>> hope,
>> having different user interface classes for dense and sparse is a bit
>> arkward
>> and it's hard to give SVC and SVR different functionality with the
>> current structure.
>>
>> I suggest putting the "nu" and "C" variants of SVC and SVR in the same
>> class,
>> as that might already make things somewhat easier.
>> What do you think about that?
>> Might hope would be to have a "BaseLibSVM" and an SVR and SVC deriving
>> from that.
>> These could then "under the hood" call dense or sparse implementation.
>>
>> Comments?
>>      
> The NuSVC and SVC might not take the same hyperparams (`C` and
> `scale_C` vs `nu`). Do you plan to keep all the hyperparam and add a
> new switch and ignore the hyperparams that are not relevant?
>
> On the plus side:
> - that would make it easier to grid search Nu-SVC vs C-SVC
>
>    
Does it ever make sense to grid search over this if you also
grid search over C/nu?

> On the minus side:
> - having a many constructor parameters with some of inactive depending
> on one another makes it more complicated for the user to understand
> the class.
>
>    
The nu-SVR takes a 'C' parameter, only the nuSVC does not.
While generally I agree that having unused parameters is a bad thing,
I think it wouldn't hurt so much here (as there are already
quite a few with the different kernels)
> I don't have any strong opinion that would favor one vs the other.
>
> In any case we should preserve backward compat with a deprecation
> warning for NuSVC if we decide to merge it into SVC.
>
>    
Definitely


Another thing is:
Does any one have a reference on a paper using nuSVMs?
I have never seen that and as such I think it is only
of interest to people studying SVM optimization.
For actual SVM users, I would guess that they
never have a need for using nuSVM and that having
all this duplication is more confusing than helping.

------------------------------------------------------------------------------
Ridiculously easy VDI. With Citrix VDI-in-a-Box, you don't need a complex
infrastructure or vast IT resources to deliver seamless, secure access to
virtual desktops. With this all-in-one solution, easily deploy virtual 
desktops for less than the cost of PCs and save 60% on VDI infrastructure 
costs. Try it free! http://p.sf.net/sfu/Citrix-VDIinabox
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general

Reply via email to