Multi-output regressor and sklearn's RFE module

Question

I was wondering if it is possible to do RFE using a multi-variate estimator with sklearn package. I checked the documentation and I can't find anything for or against it. Here is the minimal code:

import sklearn.linear_model as skl
from sklearn.feature_selection import RFE
from scat import *
from sklearn import metrics, model_selection

# -- params
n_folds = 5
N       = 5000
# -- regressor
regress = skl.RidgeCV(alphas=[1e-6, 1e-5, 1e-4, 1e-3, 1e-2, 1e-1, 1])

# -- cross-validation
P = np.random.permutation(N).reshape((n_folds, -1))
cross_val_folds = []

for i_fold in range(n_folds):
    fold = (np.concatenate(P[np.arange(n_folds) != i_fold], axis=0), P[i_fold])
    cross_val_folds.append(fold)

my_X = np.random.normal(0,1,[N, 315])
my_y = np.random.normal(0,1,[N, 2])
my_pred = model_selection.cross_val_predict(regress, X=my_X, y=my_y, cv=cross_val_folds)

MAE = metrics.mean_absolute_error(my_y, my_pred)
RMSE = np.sqrt(metrics.mean_squared_error(my_y, my_pred))
print('MAE: {}, RMSE: {}'.format(round(MAE, 5), round(RMSE, 5)))

rfe = RFE(regress, 300)
feature_final = rfe.fit_transform(my_X, my_y)

but I get the following error when testing it

ValueError: bad input shape (5000, 2)

which doesn't provide much information.

---

Edits:

Apparently, using RFE function, y goes through

y = column_or_1d(y, warn=True)

which requires y to be a vector. Is there a workaround for this?

Solution

RFE does not support multi-label format because each target would result in selection of different combination of input features. Hence, you need to create individual RFE for each target variable.

For example:

rfe = {}
for i in range(my_y.shape[1]):
    rfe[i] = RFE(regress, 300) 
    rfe[i].fit(my_X, my_y[:,i])

feature_final = rfe[0].transform(my_X)
feature_final.shape
# (5000, 300)

Note from documentation of cross_val_predict:

It is not appropriate to pass these predictions into an evaluation metric. Use cross_validate to measure generalization error.

Comments

[Kinnectus]

Additionally, by giving your computer the public IP of your connection I hope this also doesn't put that device in your modem/router DMZ and, essentially, open all ports to that device... that's just looking for trouble with a Windows box (if you haven't taken due care to manage the firewall rules)...

[Blade]

Thanks for the comment on RFE's support of multi-label format. I think creating 2 RFE's would be meaningless in this scenario. But a good idea would be doing RFE's based on different target variables in sequence.