EasyEnsemble should be a meta-estimator

[amueller]

I'm not entirely sure how EasyEnsemble should be used, but I feel like it might be easier if it was a meta-estimator.
Let's say I want to implement a random forest using EasyEnsemble as an estimator. I have no idea how to do that easily. If it was a meta-estimator, I could just do EasyEnsemble(DecisionTreeClassifier(max_features="auto")), which would be nice. Ideally this would set the random seeds, too.

I ended up with this solution, which works, but is verbose and probably inefficient:

from sklearn.ensemble import VotingClassifier
from sklearn.tree import DecisionTreeClassifier
def make_resample_tree(random_state=0):
    tree = make_imb_pipeline(RandomUnderSampler(random_state=random_state, replacement=True),
                             DecisionTreeClassifier(max_features='auto', random_state=random_state))
    return "tree_i".format(random_state), tree
classifiers = [make_resample_tree(i) for i in range(100)]
resampled_rf = VotingClassifier(classifiers, voting='soft')

[Though it is kinda neat that I can write it down like that ;]

[amueller]

A stupid implementation of the general scheme is this:

from sklearn.base import clone

def make_resampled_ensemble(estimator, n_estimators=100):
    estimators = []
    for i in range(n_estimators):
        est = clone(estimator)
        if hasattr(est, "random_state"):
            est.random_state = i
        pipe = make_imb_pipeline(RandomUnderSampler(random_state=i, replacement=True),
                                 est)
        estimators.append(("est_i".format(i), pipe))
    return VotingClassifier(estimators, voting="soft")

Clearly you could do that more elegantly by actually writing a meta-estimator class...

[glemaitre]

You are right. I think that this is also the solution at #149.

[chkoar]

@amueller is right. On the other hand we could preserve functionality by rename this class as EasyEnsembleSampler and on top of that build the EasyEnsembleClassifier. I am ok either way. I could do it. PS the sampler and the classifier could easily paralellized.

[glemaitre]

Yep exactly. We need to deprecate the class as well. I let you on this one. We should do the same for BalanceCascade G

[amueller]

fwiw I don't have a strong opinion on also keeping the current version or not - people might rely on it.

[glemaitre]

@chkoar I was thinking for that matter. I think that the resampling strategy should take any under-sampling methods and that the classifier could be any classifier as well. You should probably keep this in mind while designing the meta-estimators.