Note

Go to the end to download the full example code. or to run this example in your browser via JupyterLite or Binder

Combine predictors using stacking#

Stacking refers to a method to blend estimators. In this strategy, some estimators are individually fitted on some training data while a final estimator is trained using the stacked predictions of these base estimators.

In this example, we illustrate the use case in which different regressors are stacked together and a final linear penalized regressor is used to output the prediction. We compare the performance of each individual regressor with the stacking strategy. Stacking slightly improves the overall performance.

# Authors: The scikit-learn developers
# SPDX-License-Identifier: BSD-3-Clause

Download the dataset#

We will use the Ames Housing dataset which was first compiled by Dean De Cock and became better known after it was used in Kaggle challenge. It is a set of 1460 residential homes in Ames, Iowa, each described by 80 features. We will use it to predict the final logarithmic price of the houses. In this example we will use only 20 most interesting features chosen using GradientBoostingRegressor() and limit number of entries (here we won’t go into the details on how to select the most interesting features).

The Ames housing dataset is not shipped with scikit-learn and therefore we will fetch it from OpenML.

importnumpyasnp
fromsklearn.datasetsimport fetch_openml
fromsklearn.utilsimport shuffle
defload_ames_housing():
 df = fetch_openml (name="house_prices", as_frame=True)
 X = df.data
 y = df.target
 features = [
 "YrSold",
 "HeatingQC",
 "Street",
 "YearRemodAdd",
 "Heating",
 "MasVnrType",
 "BsmtUnfSF",
 "Foundation",
 "MasVnrArea",
 "MSSubClass",
 "ExterQual",
 "Condition2",
 "GarageCars",
 "GarageType",
 "OverallQual",
 "TotalBsmtSF",
 "BsmtFinSF1",
 "HouseStyle",
 "MiscFeature",
 "MoSold",
 ]
 X = X.loc[:, features]
 X, y = shuffle (X, y, random_state=0)
 X = X.iloc[:600]
 y = y.iloc[:600]
 return X, np.log (y)
X, y = load_ames_housing()

Make pipeline to preprocess the data#

Before we can use Ames dataset we still need to do some preprocessing. First, we will select the categorical and numerical columns of the dataset to construct the first step of the pipeline.

fromsklearn.composeimport make_column_selector
cat_selector = make_column_selector (dtype_include=object)
num_selector = make_column_selector (dtype_include=np.number )
cat_selector(X)

['HeatingQC', 'Street', 'Heating', 'MasVnrType', 'Foundation', 'ExterQual', 'Condition2', 'GarageType', 'HouseStyle', 'MiscFeature']

num_selector(X)

['YrSold', 'YearRemodAdd', 'BsmtUnfSF', 'MasVnrArea', 'MSSubClass', 'GarageCars', 'OverallQual', 'TotalBsmtSF', 'BsmtFinSF1', 'MoSold']

Then, we will need to design preprocessing pipelines which depends on the ending regressor. If the ending regressor is a linear model, one needs to one-hot encode the categories. If the ending regressor is a tree-based model an ordinal encoder will be sufficient. Besides, numerical values need to be standardized for a linear model while the raw numerical data can be treated as is by a tree-based model. However, both models need an imputer to handle missing values.

We will first design the pipeline required for the tree-based models.

fromsklearn.composeimport make_column_transformer
fromsklearn.imputeimport SimpleImputer
fromsklearn.pipelineimport make_pipeline
fromsklearn.preprocessingimport OrdinalEncoder
cat_tree_processor = OrdinalEncoder (
 handle_unknown="use_encoded_value",
 unknown_value=-1,
 encoded_missing_value=-2,
)
num_tree_processor = SimpleImputer (strategy="mean", add_indicator=True)
tree_preprocessor = make_column_transformer (
 (num_tree_processor, num_selector), (cat_tree_processor, cat_selector)
)
tree_preprocessor

ColumnTransformer(transformers=[('simpleimputer',
 SimpleImputer(add_indicator=True),
 <sklearn.compose._column_transformer.make_column_selector object at 0x7f4d15caad70>),
 ('ordinalencoder',
 OrdinalEncoder(encoded_missing_value=-2,
 handle_unknown='use_encoded_value',
 unknown_value=-1),
 <sklearn.compose._column_transformer.make_column_selector object at 0x7f4d15ca8e80>)])

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

Then, we will now define the preprocessor used when the ending regressor is a linear model.

fromsklearn.preprocessingimport OneHotEncoder , StandardScaler
cat_linear_processor = OneHotEncoder (handle_unknown="ignore")
num_linear_processor = make_pipeline (
 StandardScaler (), SimpleImputer (strategy="mean", add_indicator=True)
)
linear_preprocessor = make_column_transformer (
 (num_linear_processor, num_selector), (cat_linear_processor, cat_selector)
)
linear_preprocessor

ColumnTransformer(transformers=[('pipeline',
 Pipeline(steps=[('standardscaler',
 StandardScaler()),
 ('simpleimputer',
 SimpleImputer(add_indicator=True))]),
 <sklearn.compose._column_transformer.make_column_selector object at 0x7f4d15caad70>),
 ('onehotencoder',
 OneHotEncoder(handle_unknown='ignore'),
 <sklearn.compose._column_transformer.make_column_selector object at 0x7f4d15ca8e80>)])

Measure and plot the results#

Now we can use Ames Housing dataset to make the predictions. We check the performance of each individual predictor as well as of the stack of the regressors.

importtime
importmatplotlib.pyplotasplt
fromsklearn.metricsimport PredictionErrorDisplay
fromsklearn.model_selectionimport cross_val_predict , cross_validate
fig, axs = plt.subplots (2, 2, figsize=(9, 7))
axs = np.ravel (axs)
for ax, (name, est) in zip(
 axs, estimators + [("Stacking Regressor", stacking_regressor)]
):
 scorers = {"R2": "r2", "MAE": "neg_mean_absolute_error"}
 start_time = time.time ()
 scores = cross_validate (
 est, X, y, scoring=list(scorers.values()), n_jobs=-1, verbose=0
 )
 elapsed_time = time.time () - start_time
 y_pred = cross_val_predict (est, X, y, n_jobs=-1, verbose=0)
 scores = {
 key: (
 f"{np.abs(np.mean (scores[f'test_{value}'])):.2f} +- "
 f"{np.std (scores[f'test_{value}']):.2f}"
 )
 for key, value in scorers.items()
 }
 display = PredictionErrorDisplay.from_predictions (
 y_true=y,
 y_pred=y_pred,
 kind="actual_vs_predicted",
 ax=ax,
 scatter_kwargs={"alpha": 0.2, "color": "tab:blue"},
 line_kwargs={"color": "tab:red"},
 )
 ax.set_title(f"{name}\nEvaluation in {elapsed_time:.2f} seconds")
 for name, score in scores.items():
 ax.plot([], [], " ", label=f"{name}: {score}")
 ax.legend(loc="upper left")
plt.suptitle ("Single predictors versus stacked predictors")
plt.tight_layout ()
plt.subplots_adjust (top=0.9)
plt.show ()

Single predictors versus stacked predictors, Random Forest Evaluation in 1.07 seconds, Lasso Evaluation in 0.23 seconds, Gradient Boosting Evaluation in 0.46 seconds, Stacking Regressor Evaluation in 8.95 seconds

The stacked regressor will combine the strengths of the different regressors. However, we also see that training the stacked regressor is much more computationally expensive.

Total running time of the script: (0 minutes 21.840 seconds)

Launch binder

Launch JupyterLite

Download Jupyter notebook: plot_stack_predictors.ipynb

Download Python source code: plot_stack_predictors.py

Download zipped: plot_stack_predictors.zip

Related examples

Plot individual and voting regression predictions

Categorical Feature Support in Gradient Boosting

Displaying estimators and complex pipelines

Decision Tree Regression with AdaBoost

Gallery generated by Sphinx-Gallery