Support for MultiOutputClassifier

[martianescape]

I am trying to use SMOTE along with Randomforest multiOutputClassifier.

Here is how I am using it,

forest = RandomForestClassifier(n_estimators= n_estimator, random_state=1)
multi_target_forest = MultiOutputClassifier(forest, n_jobs=-1)
sampler = SMOTE()
clf = make_pipeline(sampler, multi_target_forest)
model = clf.fit(vecs, vec_labels)

Is there a support for multOutputClassifier? Is this the correct way to use it?

I am encountering the value error in vec_labels.
length of vecs is 1631, vec_labels is [1631][241].
Please help.

[chkoar]

I think that scikit-learn's RandomForestClassifier supports multi-output problems out of the box.
Anyway, assuming that you import the make_pipeline from the imbalanced-learn package, could you try the snippet below?

forest = RandomForestClassifier(n_estimators=n_estimator, random_state=1)
sampler = SMOTE(random_state=2)
pipeline = make_pipeline(sampler, forest)
multi_target_forest = MultiOutputClassifier(pipeline, n_jobs=-1)
model = multi_target_forest.fit(vecs, vec_labels)

[glemaitre]

If the method of @chkoar (which should work) does not work, I don't think we should support mutli-target. In some way, the bottleneck will be for theoretical reason since the algorithms were not designed to work for multi-targets.

@chkoar WDYT?

[aqibsaeed]

I also think this would be a very nice addition to support multi-class oversampling out of the box with imbalanced-learn.

[glemaitre]

I also think this would be a very nice addition to support multi-class oversampling out of the box with imbalanced-learn.

It is already supported.

[glemaitre]

Closed if favor of #340 regarding support of multilabel/multiclass