如何解决“RFECV 对象没有 support_ 属性”属性错误？

Question

我正在尝试传递 sklearn RFECV 对象并交叉验证分数以返回具有所选特征和特征排名的模型性能。

但是，我很可能收到“RFECV 对象没有 support_attribute”错误，因为我没有将它拟合到数据中。我需要一些帮助来确定适合数据的位置以及如何确保测试数据集没有数据泄漏。

原始数据集是时序数据，所以我使用 TimeSeries Split 进行拆分。

from sklearn.datasets import make_classification
from sklearn.feature_selection import RFE, RFECV
from sklearn.tree import DecisionTreeClassifier
from sklearn.model_selection import TimeSeriesSplit, cross_val_score
from sklearn import metrics
from sklearn.metrics import balanced_accuracy_score, make_scorer

X, y = make_classification(n_samples=1000, n_features=10, n_informative=5, n_redundant=5, random_state=1)

# create pipeline
rfecv_model = RFECV(estimator=DecisionTreeClassifier())
model = DecisionTreeClassifier()
pipeline = Pipeline(steps=[('s',rfecv_model),('m',model)])

#make balanced scorer
scorer = make_scorer(balanced_accuracy_score)

# evaluate model
cv = TimeSeriesSplit(n_splits=3)
n_scores = cross_val_score(pipeline, X, y, scoring=scorer, cv=cv)
# report performance
print('Balanced_Accuracy: %.3f (%.3f)' % (mean(n_scores), std(n_scores)))

for i in range(X.shape[1]):
    print('Column: %d, Selected %s, Rank: %.3f' % (i, rfecv_model.support_[i], rfecv_model.ranking_[i]))

此代码来源于RFE教程here

Answer 1

我会建议使用cross_validate当您需要拟合型号的交叉验证时。

from sklearn import set_config

set_config(print_changed_only=True)


from sklearn.datasets import make_classification
from sklearn.feature_selection import RFE, RFECV
from sklearn.tree import DecisionTreeClassifier
from sklearn.model_selection import TimeSeriesSplit, cross_validate
from sklearn import metrics
from sklearn.metrics import balanced_accuracy_score, make_scorer
from sklearn.pipeline import Pipeline

X, y = make_classification(
    n_samples=1000, n_features=10, n_informative=5, n_redundant=5, random_state=1)

# create pipeline
rfecv_model = RFECV(estimator=DecisionTreeClassifier())
model = DecisionTreeClassifier()
pipeline = Pipeline(steps=[('s', rfecv_model), ('m', model)])

# make balanced scorer
scorer = make_scorer(balanced_accuracy_score)

# evaluate model
cv = TimeSeriesSplit(n_splits=3)
result = cross_validate(pipeline, X, y, scoring=scorer,
                          cv=cv, return_estimator=True)

结果 h3>

{'fit_time': array([0.07009673, 0.09101987, 0.11680794]),
 'score_time': array([0.00072193, 0.00065613, 0.00060487]),
 'estimator': (Pipeline(steps=[('s', RFECV(estimator=DecisionTreeClassifier())),
                  ('m', DecisionTreeClassifier())]),
  Pipeline(steps=[('s', RFECV(estimator=DecisionTreeClassifier())),
                  ('m', DecisionTreeClassifier())]),
  Pipeline(steps=[('s', RFECV(estimator=DecisionTreeClassifier())),
                  ('m', DecisionTreeClassifier())])),
 'test_score': array([0.812     , 0.83170092, 0.8510502 ])}

现在让我们通过特征选择器进行CV的每次迭代。

for iter, pipe in enumerate(result['estimator']):
    print(f'Iteration no: {iter}')
    for i in range(X.shape[1]):
        print('Column: %d, Selected %s, Rank: %d' %
            (i, pipe['s'].support_[i], pipe['s'].ranking_[i]))

# output
Iteration no: 0
Column: 0, Selected False, Rank: 4
Column: 1, Selected True, Rank: 1
Column: 2, Selected True, Rank: 1
Column: 3, Selected True, Rank: 1
Column: 4, Selected False, Rank: 3
Column: 5, Selected False, Rank: 5
Column: 6, Selected True, Rank: 1
Column: 7, Selected True, Rank: 1
Column: 8, Selected True, Rank: 1
Column: 9, Selected False, Rank: 2
Iteration no: 1
Column: 0, Selected False, Rank: 2
Column: 1, Selected False, Rank: 4
Column: 2, Selected True, Rank: 1
Column: 3, Selected True, Rank: 1
Column: 4, Selected True, Rank: 1
Column: 5, Selected False, Rank: 6
Column: 6, Selected True, Rank: 1
Column: 7, Selected False, Rank: 5
Column: 8, Selected True, Rank: 1
Column: 9, Selected False, Rank: 3
Iteration no: 2
Column: 0, Selected True, Rank: 1
Column: 1, Selected False, Rank: 4
Column: 2, Selected True, Rank: 1
Column: 3, Selected True, Rank: 1
Column: 4, Selected True, Rank: 1
Column: 5, Selected False, Rank: 3
Column: 6, Selected False, Rank: 2
Column: 7, Selected False, Rank: 5
Column: 8, Selected True, Rank: 1
Column: 9, Selected True, Rank: 1