Adding feature scaling to a nested cross-validation with randomized search and recursive feature elimination

user_8689518 0
Viewed 262

I have a classification task and want to use a repeated nested cross-validation to simultaneously perform hyperparameter tuning and feature selection. For this, I am running RandomizedSearchCV on RFECV using Python's sklearn library, as suggested in this SO answer.

However, I additionally need to scale my features and impute some missing values first. Those two steps should also be included into the CV framework to avoid information leakage between training and test folds. I tried to create a Pipeline to get there but I think it "destroys" my CV-nesting (i.e., performs the RFECV and random search separately from each other):

import numpy as np    
from sklearn.datasets import make_classification
from sklearn.model_selection import RepeatedStratifiedKFold
from sklearn.feature_selection import RFECV
import scipy.stats as stats
from sklearn.utils.fixes import loguniform
from sklearn.preprocessing import StandardScaler
from sklearn.impute import KNNImputer
from sklearn.linear_model import SGDClassifier
from sklearn.model_selection import RandomizedSearchCV
from sklearn.pipeline import Pipeline

# create example data with missings
Xtrain, ytrain = make_classification(n_samples = 500,
                                     n_features = 150,
                                     n_informative = 25,
                                     n_redundant = 125,
                                     random_state = 1897)
c = 10000 # number of missings
Xtrain.ravel()[np.random.choice(Xtrain.size, c, replace = False)] = np.nan # introduce random missings

folds = 5
repeats = 5
rskfold = RepeatedStratifiedKFold(n_splits = folds, n_repeats = repeats, random_state = 1897)
n_iter = 100

scl = StandardScaler()
imp = KNNImputer(n_neighbors = 5, weights = 'uniform')
sgdc = SGDClassifier(loss = 'log', penalty = 'elasticnet', class_weight = 'balanced', random_state = 1897)
sel = RFECV(sgdc, cv = folds)
pipe = Pipeline([('scaler', scl),
                 ('imputer', imp),
                 ('selector', sel),
                 ('clf', sgdc)])
param_rand = {'clf__l1_ratio': stats.uniform(0, 1),
              'clf__alpha': loguniform(0.001, 1)}
rskfold_search = RandomizedSearchCV(pipe, param_rand, n_iter = n_iter, cv = rskfold, scoring = 'accuracy', random_state = 1897, verbose = 1, n_jobs = -1)
rskfold_search.fit(Xtrain, ytrain)

Does anyone know how to include scaling and imputation into the CV framework without losing the nesting of my RandomizedSearchCV and RFECV?

Any help is highly appreciated!

scikit-learn python
1 Answers

You haven't lost the nested cv.

You have a search object at the top level; when you call fit, it splits the data into multiple folds. Let's focus on one such train fold. Your pipeline gets fitted on that, so you scale and impute, then the RFECV gets it to split into inner folds. Finally a new estimator gets fitted on the outer training fold, and scored on the outer testing fold.

That means the RFE is getting perhaps a little leakage, since scaling and imputing happen before its splits. You can add them in a pipeline before the estimator, and use that pipeline as the RFE estimator. And since RFECV refits its estimator using the discovered optimal number of features and exposes that for predict and so on, you don't really need the second copy of sgdc; using just the one copy has the side effect of hyperparameter-tuning the selection as well:

scl = StandardScaler()
imp = KNNImputer(n_neighbors=5, weights='uniform')
sgdc = SGDClassifier(loss='log', penalty='elasticnet', class_weight='balanced', random_state=1897)
base_pipe = Pipeline([
    ('scaler', scl),
    ('imputer', imp),
    ('clf', sgdc),
])
sel = RFECV(base_pipe, cv=folds)

param_rand = {'estimator__clf__l1_ratio': stats.uniform(0, 1),
              'estimator__clf__alpha': loguniform(0.001, 1)}
rskfold_search = RandomizedSearchCV(sel, param_rand, n_iter=n_iter, cv=rskfold, scoring='accuracy', random_state=1897, verbose=1, n_jobs=-1)
user_104958930
Related