Scikit-Learn RandomizedSearchCV StackingClassifier

Hyperparameter tuning is essential for optimizing machine learning models. In this example, we’ll demonstrate how to use scikit-learn’s RandomizedSearchCV for hyperparameter tuning of a StackingClassifier, which combines multiple models to improve predictive performance.

Random search is a method for evaluating different combinations of model hyperparameters. Unlike grid search, it samples a fixed number of hyperparameter combinations from a specified distribution, making it more efficient when searching over a large hyperparameter space.

StackingClassifier is a meta-estimator that combines the predictions of several base estimators using another estimator as the final classifier. This approach leverages the strengths of multiple models to achieve better performance.

Key hyperparameters for StackingClassifier include the choice of base estimators, the final estimator, and the specific hyperparameters of each estimator. These parameters significantly impact the model’s performance and need to be carefully tuned.

from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split, RandomizedSearchCV
from sklearn.ensemble import StackingClassifier
from sklearn.linear_model import LogisticRegression
from sklearn.tree import DecisionTreeClassifier
from sklearn.neighbors import KNeighborsClassifier
from scipy.stats import uniform, randint

# Generate synthetic binary classification dataset
X, y = make_classification(n_samples=100, n_features=20, n_informative=10, n_redundant=10, random_state=42)

# Split into train and test sets
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

# Define base estimators
estimators = [
    ('lr', LogisticRegression(random_state=42)),
    ('dt', DecisionTreeClassifier(random_state=42)),
    ('knn', KNeighborsClassifier())
]

# Define the model
model = StackingClassifier(estimators=estimators, final_estimator=LogisticRegression())

# Define hyperparameter distribution
param_dist = {
    'lr__C': uniform(loc=0, scale=4),
    'dt__max_depth': randint(1, 20),
    'knn__n_neighbors': randint(1, 20),
    'final_estimator__C': uniform(loc=0, scale=4)
}

# Perform random search
random_search = RandomizedSearchCV(estimator=model,
                                   param_distributions=param_dist,
                                   n_iter=50,
                                   cv=5,
                                   scoring='accuracy',
                                   random_state=42)
random_search.fit(X_train, y_train)

# Report best score and parameters
print(f"Best score: {random_search.best_score_:.3f}")
print(f"Best parameters: {random_search.best_params_}")

# Evaluate on test set
best_model = random_search.best_estimator_
accuracy = best_model.score(X_test, y_test)
print(f"Test set accuracy: {accuracy:.3f}")

Running the example gives an output like:

Best score: 0.938
Best parameters: {'dt__max_depth': 15, 'final_estimator__C': 2.473544037332349, 'knn__n_neighbors': 12, 'lr__C': 2.0569377536544464}
Test set accuracy: 0.750

The steps are as follows:

1. Generate a synthetic binary classification dataset using scikit-learn’s make_classification function.
2. Split the dataset into train and test sets using train_test_split.
3. Define the StackingClassifier with base estimators (LogisticRegression, DecisionTreeClassifier, KNeighborsClassifier) and a final estimator (LogisticRegression).
4. Define hyperparameter distributions for the base and final estimators, including ranges for C, max_depth, and n_neighbors.
5. Perform random search using RandomizedSearchCV, specifying the StackingClassifier, hyperparameter distribution, 100 iterations, 5-fold cross-validation, and accuracy scoring metric.
6. Report the best cross-validation score and best set of hyperparameters found by random search.
7. Evaluate the best model on the hold-out test set and report the accuracy.

By using RandomizedSearchCV, we can efficiently explore different hyperparameter settings and find the combination that maximizes the model’s performance. This automated approach saves time and effort compared to manual hyperparameter tuning and helps ensure we select the best configuration for our StackingClassifier.