[scikit-learn] Three new scikit-learn-contrib projects

[Andreas Mueller]

On 07/20/2016 01:31 PM, Guillaume Lemaître wrote:
> Hi Gael,
>
> I was wondering if you could elaborate on the problem of 
> hyper-parameter tuning and why the imbalanced-learn would not benefit 
> from it.
> Since that we used the identical pipeline of scikit-learn and add the 
> part to handle the sampler, I would have think that we could use it.
>
> However this is true that I did not play to much with this part of the 
> API, so I should probably missed something.
>
The assumption is that hyper-parameter tuning uses Pipelines, I think.
You want to select all steps in your processing, which is rarely just a 
single model.

However, Pipeline can currently not change the number of samples (see 
the enhancement proposal Gael linked to).
So you can not use your methods in the standard scikit-learn pipeline.

Best,
Andy