GridsearchSV 可以包含 train_test_split 的随机化吗

Question

With Sklearn there is GridSearchCV to test multiple variables for classifier functions, for example:

parameters = {
  'learning_rate': [0.001,0.005,0.003],
  'n_estimators': [300,800,1200],
  'criterion': ['friedman_mse','mse','mae'],
  'verbose':[1],
  'loss' : ['deviance','exponential'],
  'random_state':[0]
  }

GBC = GradientBoostingClassifier()
grid = GridSearchCV(GBC, parameters)
grid.fit(X,y )   # X = data,  y = result
best_est = grid.best_estimator_
print(best_est)

predictions = best_est.predict(T) # T contains data to apply it on.

But what if one would like to do cross validation? Eg in a similar manner to train_test_split :

  X_train, X_test, y_train, y_test = train_test_split(X, y,  random_state=41)

Here we have a random_state (which might have a big impact). Is it possible to include the GridSearchCV with an array of a few random numbers to make sure it works optimally with 'most' random state's of some data's train/test split?

For the record, I know this isn't inside GridSearchCV (or as far as i know), I'm asking here what such a method might look like. Perhaps there is some clever way to do this?

Answer 1

You can specify ShuffleSplit as a cross-validation generator.

For example:

from sklearn.model_selection import GridSearchCV, ShuffleSplit

GBC = GradientBoostingClassifier()
grid = GridSearchCV(GBC,
                    param_grid=parameters,
                    cv=ShuffleSplit(train_size=X.shape[0],
                                    test_size=.3,
                                    n_splits=5,
                                    random_state=41))
grid.fit(X, y)

More on ShuffleSplit here .