我叫 Lucas，对机器学习领域比较陌生。我在一些在线文档和教程的帮助下编写了这段代码。但是，我需要一些帮助来了解 RFE() 与 GBM() 的集成是否正确。
def evaluateAlgorithm(X_train, X_test, y_train, y_test, dataset):
Kfold = StratifiedKFold(n_splits=20, shuffle=True)
GBM = GradientBoostingClassifier(
loss='log_loss', learning_rate=0.01,
n_estimators=1000, subsample=0.9,
min_samples_split=2, min_samples_leaf=1,
min_weight_fraction_leaf=0.0, max_depth=8,
init=None, random_state=None,
max_features=None, verbose=0,
max_leaf_nodes=None, warm_start=False)
pipeline = Pipeline(steps=[['feature_selection', RFE(GBM)], ['model', GBM]])
parameters = {'model__learning_rate': [0.01, 0.02, 0.03],
'model__subsample': [0.9, 0.5, 0.3, 0.1],
'model__n_estimators': [100, 500, 1000],
'model__max_depth': [1, 2, 3],
'feature_selection__n_features_to_select': [7, 14, 27]}
grid_GBM = GridSearchCV(estimator=pipeline, param_grid=parameters, cv=Kfold,
verbose=1, n_jobs=-1, refit=True, scoring='accuracy')
grid_GBM.fit(X_train, y_train)
print("\n=========================================================================")
print(" Results from Grid Search Gradient Boosting")
print("=========================================================================")
print("\n The best estimator across ALL searched params: \n", grid_GBM.best_estimator_)
print("\n The best score across ALL searched params: \n", grid_GBM.best_score_)
print("\n The best parameters across ALL searched params: \n", grid_GBM.best_params_)
print("\n=========================================================================")
# Obtain features selected by RFE
rfe_selected_features_indices = grid_GBM.best_estimator_['feature_selection'].support_
rfe_selected_features_names = X_train.columns[rfe_selected_features_indices]
print("Features selected by RFE:", rfe_selected_features_names)
model_GBM = grid_GBM.best_estimator_
# Cross-validation
cv_results_GBM = cross_val_score(model_GBM, X_train, y_train, cv=Kfold, scoring='accuracy', n_jobs=-1, verbose=0)
print()
print("Cross Validation results Gradient Boosting: ", cv_results_GBM)
prt_string = "CV Mean accuracy: %f (Std: %f)" % (cv_results_GBM.mean(), cv_results_GBM.std())
print(prt_string)
trained_Model_GBM = model_GBM.fit(X_train, y_train)
print();
print('========================================================')
print();
print(trained_Model_GBM.get_params(deep=True))
print();
print('=========================================================')
# Make predictions on the test set
pred_Labels_GBM = trained_Model_GBM.predict(X_test)
pred_proba_GBM = trained_Model_GBM.predict_proba(X_test)
# Evaluate performance
print();
print('Evaluation of the trained model Gradient Boosting: ')
accuracy = accuracy_score(y_test, pred_Labels_GBM)
print();
print('Accuracy Gradient Boosting: ', accuracy)
precision = precision_score(y_test, pred_Labels_GBM, pos_label='positive')
print();
print('Precision Gradient Boosting: ', precision)
recall = recall_score(y_test, pred_Labels_GBM, pos_label='positive')
print();
print('Recall Score Gradient Boosting: ', recall)
f1 = f1_score(y_test, pred_Labels_GBM, pos_label='positive')
print();
print('f1 Score Gradient Boosting: ', f1)
confusion_mat = confusion_matrix(y_test, pred_Labels_GBM)
classReport = classification_report(y_test, pred_Labels_GBM)
print();
print('Classification Report Gradient Boosting: \n', classReport)
kappa_score = cohen_kappa_score(y_test, pred_Labels_GBM)
print();
print('Kappa Score Gradient Boosting: ', kappa_score)
skplt.estimators.plot_learning_curve(model_GBM, X_train, y_train, figsize=(8, 6))
plt.show()
skplt.metrics.plot_roc(y_test, pred_proba_GBM, figsize=(8, 6));
plt.show()
skplt.metrics.plot_confusion_matrix(y_test, pred_Labels_GBM, figsize=(8, 6));
plt.show()
skplt.metrics.plot_precision_recall(y_test, pred_proba_GBM,
title='Precision-Recall Curve', plot_micro=True,
classes_to_plot=None, ax=None, figsize=(8, 6),
cmap='nipy_spectral', title_fontsize='large',
text_fontsize='medium');
plt.show()
evaluateAlgorithm(X_train, X_test, y_train, y_test, dataset)
我的目标是使用 RFE 找到最佳的特征组合，同时使用 GBM 的网格搜索来寻找最佳的超参数。然而，RFE 似乎只是在网格搜索的超参数之前找到最好的特征。我该如何解决这个问题以便两个过程同时发生？我们的想法是实现这两个标准的最佳组合。此外，您对改进此代码有什么建议吗？
如所写，您的代码调整最终模型的超参数，而不是特征选择步骤中 GBM 的超参数。有几个选择：
扩展搜索空间以包含选择GBM的超参数，例如
feature_selection__estimator__max_depth删除模型步骤。
RFEestimator_RFErfe.predict这些方法之间的区别在于，第一种方法允许选择 GBM 具有与模型 GBM 不同的超参数。这往往会更昂贵，但更灵活。如果它提供了显着的改进，我个人会感到惊讶，所以我建议第二种方法，除非您有时间并且愿意进行实验。