Allow RandomUp/DownSampler to run with Non-numeric columns for Pre-Encoding Sampling

[JoshuaC3]

I wished to use undersampling before doing categorical encoding. Naturally, I have strings in the columns so the .fit_sample or simply .fit methods throw errors because it cannot use X with non-numerics.

I think there should be a kwarg that can be passed so that the checks are not made. I can then encode my categoricals (using stats based encoding such as count or mean based encoding) afterwards.

[glemaitre]

I don't think kwargs is the solution. However, both RandomOverSampler and RandomUnderSampler could lift the numeric enforcement as you mentioned. We might want to make the checking into a private function _check_inputs. Like this we can redefine it in RandomOverSampler and RandomUnderSampler.

[JoshuaC3]

Ok - ultimately, anything that gives a little extra flexibility to allow pre-categorical-encoding up/down sampling would be a great thing.

Do you mean make _check_inputs a private function and simply not call it on RandomUnderSampler or pass a kwarg to turn it on/off? Would be happy to do a PR but not sure exactly what is wanted. Thanks.

[glemaitre]

Do you mean make _check_inputs a private function and simply not call it on RandomUnderSampler or pass a kwarg to turn it on/off? Would be happy to do a PR but not sure exactly what is wanted. Thanks.

I would say to create a _check_inputs in the base class which is overwritten to be more permissive in RandomUnderSampling

class BaseSampler(...):
    def _check_inputs(self, X, y):
        ...
        X, y = check_array(X, y)
        ...

class RandomUnderSampler(...):
    def _check_inputs():
        ....
        #overwrite the call to check_array
        X, y = check_array(X, y, dtype=None)

Something similar to that.

[glemaitre]

This is addressed in #429 and should be merged when the CI are green