Why do I get different values with pipline and without pipline in sklearn in python

Question

I am using recursive feature elimination with cross-validation (rfecv) with GridSearchCV with RandomForest classifier as follows using pipeline and without using pipeline.

My code with pipeline is as follows.

X = df[my_features_all]
y = df['gold_standard']

#get development and testing sets
x_train, x_test, y_train, y_test = train_test_split(X, y, random_state=0)

from sklearn.pipeline import Pipeline

#cross validation setting
k_fold = StratifiedKFold(n_splits=5, shuffle=True, random_state=0)
#this is the classifier used for feature selection
clf_featr_sele = RandomForestClassifier(random_state = 42, class_weight="balanced")
rfecv = RFECV(estimator=clf_featr_sele, step=1, cv=k_fold, scoring='roc_auc')

param_grid = {'n_estimators': [200, 500],
    'max_features': ['auto', 'sqrt', 'log2'],
    'max_depth' : [3,4,5]
    }

#you can have different classifier for your final classifier
clf = RandomForestClassifier(random_state = 42, class_weight="balanced")
CV_rfc = GridSearchCV(estimator=clf, param_grid=param_grid, cv= k_fold, scoring = 'roc_auc', verbose=10, n_jobs = 5)

pipeline  = Pipeline([('feature_sele',rfecv),('clf_cv',CV_rfc)])

pipeline.fit(x_train, y_train)

The result is (with pipeline):

Optimal features: 29
Best hyperparameters: {'max_depth': 3, 'max_features': 'auto', 'n_estimators': 500}
Best score: 0.714763

My code without pipeline is as follows.

X = df[my_features_all]
y = df['gold_standard']

#get development and testing sets
x_train, x_test, y_train, y_test = train_test_split(X, y, random_state=0)

#cross validation setting
k_fold = StratifiedKFold(n_splits=5, shuffle=True, random_state=0)

clf = RandomForestClassifier(random_state = 42, class_weight="balanced")

rfecv = RFECV(estimator=clf, step=1, cv=k_fold, scoring='roc_auc')

param_grid = {'estimator__n_estimators': [200, 500],
    'estimator__max_features': ['auto', 'sqrt', 'log2'],
    'estimator__max_depth' : [3,4,5]
    }

CV_rfc = GridSearchCV(estimator=rfecv, param_grid=param_grid, cv= k_fold, scoring = 'roc_auc', verbose=10, n_jobs = 5)
CV_rfc.fit(x_train, y_train)

The result is (without pipeline):

Optimal features: 4
Best hyperparameters: {'max_depth': 3, 'max_features': 'auto', 'n_estimators': 500}
Best score: 0.756835

Even though, the concept of both approaches is similar I get different results and different selected features (as shown above in results sections). However, I get the same hyperparameter values.

I am just wondering why this difference happens. What approach (without using pipeline or with using pipeline?) is the most suitable to perform the aforementioned task?

I am happy to provide more details if needed.

Answer 1

In with pipeline case,

Feature selection (RFECV) is carried out with base model (RandomForestClassifier(random_state = 42, class_weight="balanced")) before applying the grid_searchCV on final estimator.

In without pipeline case,

For each combination of hyperparameter, the corresponding estimator is used for feature selection (RFECV). Hence, it would be time consuming.