Enabled feature_importances_ for our ForestDML and ForestDRLearner estimators

[vsyrgkanis]

This required changing the subsampled honest forest code a bit so that it does not alter the arrays of the tree structures of sklearn but rather stores two additional arrays required for prediction. This does add around 1.5 times the original running time, so makes it slightly slower due to the extra memory allocation.

However this enables correct feature_importance calculation and also in the future correct SHAP calculation (fixes #297), as now the tree entries are consistent with a tree in a randomforestregressor and so shap logic can be applied if we recast the subsampled honest forest as a randomforestregressor (additivity of shap will still be violated since the prediction of the subsample honest forest is not just the aggregation of the predictions across the trees but more complex weighted average). But we can still call shap and still get meaningful shap numbers. One discrepancy is that shap is explaining a different value that what effect returns, since it explains the value that corresponds to the average of the predictions of each honest tree regressor. however, the prediction of an honest forest is not the average of the tree predictions. For a full solution to this small discrepancy, one would need a full re-working of Shap's tree explainer and the tree explainer algorithm to account for such alternative aggregations of tree predictors.

This enables the following two uses:

from econml.dml import ForestDMLCateEstimator
import shap
import sklearn.ensemble
import copy
est3 = ForestDMLCateEstimator(model_y=RandomForestRegressor(),
                              model_t=RandomForestRegressor(),
                              n_estimators=1000,
                              subsample_fr=.8,
                              min_samples_leaf=10,
                              min_impurity_decrease=0.001,
                              verbose=0, min_weight_fraction_leaf=.01)
est3.fit(Y, T, X, W)
print(est3.feature_importances_)
model = copy.deepcopy(est3.model_cate)
model.__class__ = sklearn.ensemble.RandomForestRegressor
explainer = shap.Explainer(model, X)
shap_values = explainer(X)
shap.plots.beeswarm(shap_values, X)

from econml.drlearner import ForestDRLearner
import shap
import sklearn.ensemble
import copy
est3 = ForestDRLearner(model_regression=RandomForestRegressor(),
                       model_propensity=RandomForestClassifier(min_samples_leaf=10),
                       min_propensity=1e-3,
                       n_estimators=1000,
                       subsample_fr=.8,
                       min_samples_leaf=10,
                       min_impurity_decrease=0.001,
                       verbose=0, min_weight_fraction_leaf=.01)
est3.fit(Y, T, X, W)
for t in np.unique(T):
    if t > 0:
        print(est3.feature_importances_(T=t))
        model = copy.deepcopy(est3.model_cate(T=t))
        model.__class__ = sklearn.ensemble.RandomForestRegressor
        explainer = shap.Explainer(model, X)
        shap_values = explainer(X)
        shap.plots.beeswarm(shap_values, X)

Side benefits:

1. 6x speed up of SubsampledHonestForest by pre-transforming dense but sparsely represented matrices to dense representation before a loop with many slicing operations on them.

[kbattocchi]

Mostly looks fine, though I'm not really familiar with the tree internals to vouch for correctness. Please consider addressing the comments I left before merging.