What is _passthrough_scorer and How Can I Change Scorers in GridsearchCV (sklearn)?

Question

http://scikit-learn.org/stable/modules/generated/sklearn.grid_search.GridSearchCV.html (for reference)

x = [[2], [1], [3], [1] ... ] # about 1000 data 
grid = GridSearchCV(KernelDensity(), {'bandwidth': np.linspace(0.1, 1.0, 10)}, cv=10)
grid.fit(x)

When I use GridSearchCV without specifying scoring function like the , the value of grid.scorer_ is . Could you explain what kind of function _passthrough_scorer is?

In addition to this, I want to change the scoring function to mean_squared_error or something else.

grid = GridSearchCV(KernelDensity(), {'bandwidth': np.linspace(0.1, 1.0, 10)}, cv=10, scoring='mean_squared_error')

But the line, grid.fit(x), always gives me this error message:

TypeError: __call__() missing 1 required positional argument: 'y_true'

I cannot figure out how to give y_true to the function because I do not know the true distribution. Would you tell me how to change scoring functions? I appreciate your help.

Answer 1

The default metric for KernelDensity is minkowski with p=2 which is a a euclidean metric. GridSearchCV will use KernelDensity metric for scoring if you do not assign any other scoring method.

The formula for mean squared error is: sum((y_true - y_estimated)^2)/n. You got the error since you need to have a y_true to calculate it.

Here is a made-up example of applying GridSearchCV to KernelDensity :

from sklearn.neighbors import KernelDensity
from sklearn.grid_search import GridSearchCV
import numpy as np

N = 20
X = np.concatenate((np.random.randint(0, 10, 50),
                    np.random.randint(5, 10, 50)))[:, np.newaxis]

params = {'bandwidth': np.logspace(-1.0, 1.0, 10)}
grid = GridSearchCV(KernelDensity(), params)
grid.fit(X)
print(grid.grid_scores_)
print('Best parameter: ',grid.best_params_)
print('Best score: ',grid.best_score_)
print('Best estimator: ',grid.best_estimator_)

and output is:

[mean: -96.94890, std: 100.60046, params: {'bandwidth': 0.10000000000000001},


 mean: -70.44643, std: 40.44537, params: {'bandwidth': 0.16681005372000587},
 mean: -71.75293, std: 18.97729, params: {'bandwidth': 0.27825594022071243},
 mean: -77.83446, std: 11.24102, params: {'bandwidth': 0.46415888336127786},
 mean: -78.65182, std: 8.72507, params: {'bandwidth': 0.774263682681127},
 mean: -79.78828, std: 6.98582, params: {'bandwidth': 1.2915496650148841},
 mean: -81.65532, std: 4.77806, params: {'bandwidth': 2.1544346900318834},
 mean: -86.27481, std: 2.71635, params: {'bandwidth': 3.5938136638046259},
 mean: -95.86093, std: 1.84887, params: {'bandwidth': 5.9948425031894086},
 mean: -109.52306, std: 1.71232, params: {'bandwidth': 10.0}]
 Best parameter:  {'bandwidth': 0.16681005372000587}
 Best score:  -70.4464315885
 Best estimator:  KernelDensity(algorithm='auto', atol=0, bandwidth=0.16681005372000587,
       breadth_first=True, kernel='gaussian', leaf_size=40,
       metric='euclidean', metric_params=None, rtol=0)

The valid scoring methods for GridSeachCV usually need y_true. In your case, you may want to change the metric of sklearn.KernelDensity to other metrics (for instance to sklearn.metrics.pairwise.pairwise_kernels, sklearn.metrics.pairwise.pairwise_distances) as grid search will use them for scoring.