

In [128]:

    
from sklearn.ensemble import RandomForestRegressor, RandomForestClassifier
from sklearn.svm import SVC
from sklearn.linear_model import LogisticRegression
from sklearn import partial_dependence
import numpy as np
from scipy.stats import norm
X = np.random.normal(0, 5, size= (1000 , 3))

B= np.array([1.22, -6.2, 0.2])
y = np.dot(X, B) + norm(0, 5).rvs(1000)
y_binary = y > 0.


rf_r_model = RandomForestRegressor()
rf_r_model.fit(X, y)

svm_model = SVC()
svm_model.fit(X, y_binary)

logit_model = LogisticRegression()
logit_model.fit(X, y_binary)









    Out[128]:





LogisticRegression(C=1.0, class_weight=None, dual=False, fit_intercept=True,
          intercept_scaling=1, max_iter=100, multi_class='ovr', n_jobs=1,
          penalty='l2', random_state=None, solver='liblinear', tol=0.0001,
          verbose=0, warm_start=False)

Plug in dummy classifier :

Works with a hack



In [97]:

    
def f(array):
    return (np.sum(X, axis=1) > 0.1).astype(float)


from sklearn.dummy import DummyRegressor
c = DummyRegressor()
c.fit(X, y)
c.predict = f

pdp, axes = partial_dependence.partial_dependence(c, [0, 1], X = X, method='exact')

Use on classification model with exact method:

Fails due to bad error handling



In [104]:

    
pdp, axes = partial_dependence.partial_dependence(logit_model, [0, 1], X = X, method='exact')









    



---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-104-65649354d42c> in <module>()
----> 1 pdp, axes = partial_dependence.partial_dependence(logit_model, [0, 1], X = X, method='exact')

/usr/local/lib/python2.7/dist-packages/sklearn/partial_dependence.pyc in partial_dependence(est, target_variables, grid, X, output, percentiles, grid_resolution, method)
    307         raise ValueError('est must be a fitted regressor or classifier model.')
    308     if method != 'recursion' and est._estimator_type == 'classifier':
--> 309         raise ValueError('est requires a predict_proba method for '
    310                          'method="exact" or "estimated" for classification.')
    311     if method == 'recursion':

ValueError: est requires a predict_proba method for method="exact" or "estimated" for classification.



In [109]:

    
pdp, axes = partial_dependence.partial_dependence(logit_model, [0, 1], X = X, method='estimated')









    



---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-109-f5d4ec14fae1> in <module>()
----> 1 pdp, axes = partial_dependence.partial_dependence(logit_model.predict_proba, [0, 1], X = X, method='estimated')

/usr/local/lib/python2.7/dist-packages/sklearn/partial_dependence.pyc in partial_dependence(est, target_variables, grid, X, output, percentiles, grid_resolution, method)
    305     if (not hasattr(est, '_estimator_type') or
    306             est._estimator_type not in ('classifier', 'regressor')):
--> 307         raise ValueError('est must be a fitted regressor or classifier model.')
    308     if method != 'recursion' and est._estimator_type == 'classifier':
    309         raise ValueError('est requires a predict_proba method for '

ValueError: est must be a fitted regressor or classifier model.

Forest regressor model with estimated, exact, and recursive, using data

Works, though estimated is pretty bad



In [124]:

    
pdp_exact, axes_exact = partial_dependence.partial_dependence(rf_r_model, [0], X = X, method='exact')
pdp_est, axes_est = partial_dependence.partial_dependence(rf_r_model, [0], X = X, method='estimated')
pdp_rec, axes_rec = partial_dependence.partial_dependence(rf_r_model, [0], X = X, method='recursion')

ax = pd.DataFrame(pdp_exact[0], index = axes_exact[0], columns=['exact']).plot()
pd.DataFrame(pdp_est[0], index = axes_est[0], columns=['estimated']).plot(ax=ax)
pd.DataFrame(pdp_rec[0], index = axes_rec[0], columns=['recursion']).plot(ax=ax)









    Out[124]:





<matplotlib.axes._subplots.AxesSubplot at 0x7fb385f232d0>

Forest regressor model with estimated, exact, and recursive, using grid and data:

Fails due requirement of either X or grid but not both



In [127]:

    
grid = [i / 10. for i in range(-70, 90)]
pdp_exact, axes_exact = partial_dependence.partial_dependence(rf_r_model, [0], grid = grid, X=X, method='exact')
pdp_est, axes_est = partial_dependence.partial_dependence(rf_r_model, [0], grid = grid, X=X, method='estimated')
pdp_rec, axes_rec = partial_dependence.partial_dependence(rf_r_model, [0], grid = grid, X=X,  method='recursion')

ax = pd.DataFrame(pdp_exact[0], index = axes_exact[0], columns=['exact']).plot()
pd.DataFrame(pdp_est[0], index = axes_est[0], columns=['estimated']).plot(ax=ax)
pd.DataFrame(pdp_rec[0], index = axes_rec[0], columns=['recursion']).plot(ax=ax)









    



---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-127-275fe1eb0c4c> in <module>()
      1 grid = [i / 10. for i in range(-70, 90)]
----> 2 pdp_exact, axes_exact = partial_dependence.partial_dependence(rf_r_model, [0], grid = grid, X=X, method='exact')
      3 pdp_est, axes_est = partial_dependence.partial_dependence(rf_r_model, [0], grid = grid, X=X, method='estimated')
      4 pdp_rec, axes_rec = partial_dependence.partial_dependence(rf_r_model, [0], grid = grid, X=X,  method='recursion')
      5 

/usr/local/lib/python2.7/dist-packages/sklearn/partial_dependence.pyc in partial_dependence(est, target_variables, grid, X, output, percentiles, grid_resolution, method)
    322         n_features = X.shape[1]
    323     if (grid is None and X is None) or (grid is not None and X is not None):
--> 324         raise ValueError('Either grid or X must be specified')
    325 
    326     target_variables = np.asarray(target_variables, dtype=np.int32,

ValueError: Either grid or X must be specified

Multiclass classification:

Fails due to bad error handling



In [132]:

    
def get_multiclass(x, Y):
    if x > np.percentile(Y, .66):
        return 1
    elif x < np.percentile(Y, .33):
        return -1
    else:
        return 0
y_multiclass = np.array(map(lambda x: get_multiclass(x, y), y))
rf_multi_model = RandomForestClassifier()
rf_multi_model.fit(X, y_multiclass)

pdp_exact, axes_exact = partial_dependence.partial_dependence(rf_multi_model, [0], X=X, method='exact')









    



---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-132-8c5f7e1aa75b> in <module>()
     10 rf_multi_model.fit(X, y_multiclass)
     11 
---> 12 pdp_exact, axes_exact = partial_dependence.partial_dependence(rf_multi_model, [0], X=X, method='exact')

/usr/local/lib/python2.7/dist-packages/sklearn/partial_dependence.pyc in partial_dependence(est, target_variables, grid, X, output, percentiles, grid_resolution, method)
    307         raise ValueError('est must be a fitted regressor or classifier model.')
    308     if method != 'recursion' and est._estimator_type == 'classifier':
--> 309         raise ValueError('est requires a predict_proba method for '
    310                          'method="exact" or "estimated" for classification.')
    311     if method == 'recursion':

ValueError: est requires a predict_proba method for method="exact" or "estimated" for classification.



In [ ]:



In [ ]: