Column Transformer with Mixed Types

This example illustrates how to apply different preprocessing and feature extraction pipelines to different subsets of features, using ColumnTransformer. This is particularly handy for the case of datasets that contain heterogeneous data types, since we may want to scale the numeric features and one-hot encode the categorical ones.

In this example, the numeric data is standard-scaled after mean-imputation. The categorical data is one-hot encoded via OneHotEncoder, which creates a new category for missing values.

In addition, we show two different ways to dispatch the columns to the particular pre-processor: by column names and by column data types.

Finally, the preprocessing pipeline is integrated in a full prediction pipeline using Pipeline, together with a simple classification model.

# Author: Pedro Morales <part.morales@gmail.com>
#
# License: BSD 3 clause
import numpy as np

from sklearn.compose import ColumnTransformer
from sklearn.datasets import fetch_openml
from sklearn.pipeline import Pipeline
from sklearn.impute import SimpleImputer
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split, GridSearchCV

np.random.seed(0)

Load data from https://www.openml.org/d/40945

X, y = fetch_openml(
    "titanic", version=1, as_frame=True, return_X_y=True, parser="pandas"
)

# Alternatively X and y can be obtained directly from the frame attribute:
# X = titanic.frame.drop('survived', axis=1)
# y = titanic.frame['survived']

Use ColumnTransformer by selecting column by names

We will train our classifier with the following features:

Numeric Features:

  • age: float;

  • fare: float.

Categorical Features:

  • embarked: categories encoded as strings {'C', 'S', 'Q'};

  • sex: categories encoded as strings {'female', 'male'};

  • pclass: ordinal integers {1, 2, 3}.

We create the preprocessing pipelines for both numeric and categorical data. Note that pclass could either be treated as a categorical or numeric feature.

numeric_features = ["age", "fare"]
numeric_transformer = Pipeline(
    steps=[("imputer", SimpleImputer(strategy="median")), ("scaler", StandardScaler())]
)

categorical_features = ["embarked", "sex", "pclass"]
categorical_transformer = OneHotEncoder(handle_unknown="ignore")

preprocessor = ColumnTransformer(
    transformers=[
        ("num", numeric_transformer, numeric_features),
        ("cat", categorical_transformer, categorical_features),
    ]
)

Append classifier to preprocessing pipeline. Now we have a full prediction pipeline.

clf = Pipeline(
    steps=[("preprocessor", preprocessor), ("classifier", LogisticRegression())]
)

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=0)

clf.fit(X_train, y_train)
print("model score: %.3f" % clf.score(X_test, y_test))
model score: 0.790

HTML representation of Pipeline (display diagram)

When the Pipeline is printed out in a jupyter notebook an HTML representation of the estimator is displayed:

clf
Pipeline(steps=[('preprocessor',
                 ColumnTransformer(transformers=[('num',
                                                  Pipeline(steps=[('imputer',
                                                                   SimpleImputer(strategy='median')),
                                                                  ('scaler',
                                                                   StandardScaler())]),
                                                  ['age', 'fare']),
                                                 ('cat',
                                                  OneHotEncoder(handle_unknown='ignore'),
                                                  ['embarked', 'sex',
                                                   'pclass'])])),
                ('classifier', LogisticRegression())])
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.


Use ColumnTransformer by selecting column by data types

When dealing with a cleaned dataset, the preprocessing can be automatic by using the data types of the column to decide whether to treat a column as a numerical or categorical feature. sklearn.compose.make_column_selector gives this possibility. First, let’s only select a subset of columns to simplify our example.

subset_feature = ["embarked", "sex", "pclass", "age", "fare"]
X_train, X_test = X_train[subset_feature], X_test[subset_feature]

Then, we introspect the information regarding each column data type.

X_train.info()
<class 'pandas.core.frame.DataFrame'>
Int64Index: 1047 entries, 1118 to 684
Data columns (total 5 columns):
 #   Column    Non-Null Count  Dtype
---  ------    --------------  -----
 0   embarked  1045 non-null   category
 1   sex       1047 non-null   category
 2   pclass    1047 non-null   int64
 3   age       841 non-null    float64
 4   fare      1046 non-null   float64
dtypes: category(2), float64(2), int64(1)
memory usage: 35.0 KB

We can observe that the embarked and sex columns were tagged as category columns when loading the data with fetch_openml. Therefore, we can use this information to dispatch the categorical columns to the categorical_transformer and the remaining columns to the numerical_transformer.

Note

In practice, you will have to handle yourself the column data type. If you want some columns to be considered as category, you will have to convert them into categorical columns. If you are using pandas, you can refer to their documentation regarding Categorical data.

from sklearn.compose import make_column_selector as selector

preprocessor = ColumnTransformer(
    transformers=[
        ("num", numeric_transformer, selector(dtype_exclude="category")),
        ("cat", categorical_transformer, selector(dtype_include="category")),
    ]
)
clf = Pipeline(
    steps=[("preprocessor", preprocessor), ("classifier", LogisticRegression())]
)


clf.fit(X_train, y_train)
print("model score: %.3f" % clf.score(X_test, y_test))
clf
model score: 0.794
Pipeline(steps=[('preprocessor',
                 ColumnTransformer(transformers=[('num',
                                                  Pipeline(steps=[('imputer',
                                                                   SimpleImputer(strategy='median')),
                                                                  ('scaler',
                                                                   StandardScaler())]),
                                                  <sklearn.compose._column_transformer.make_column_selector object at 0x7f8f36bdca90>),
                                                 ('cat',
                                                  OneHotEncoder(handle_unknown='ignore'),
                                                  <sklearn.compose._column_transformer.make_column_selector object at 0x7f8f36bdcb80>)])),
                ('classifier', LogisticRegression())])
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.


The resulting score is not exactly the same as the one from the previous pipeline because the dtype-based selector treats the pclass column as a numeric feature instead of a categorical feature as previously:

selector(dtype_exclude="category")(X_train)
['pclass', 'age', 'fare']
selector(dtype_include="category")(X_train)
['embarked', 'sex']

Using the prediction pipeline in a grid search

Grid search can also be performed on the different preprocessing steps defined in the ColumnTransformer object, together with the classifier’s hyperparameters as part of the Pipeline. We will search for both the imputer strategy of the numeric preprocessing and the regularization parameter of the logistic regression using GridSearchCV.

param_grid = {
    "preprocessor__num__imputer__strategy": ["mean", "median"],
    "classifier__C": [0.1, 1.0, 10, 100],
}

grid_search = GridSearchCV(clf, param_grid, cv=10)
grid_search
GridSearchCV(cv=10,
             estimator=Pipeline(steps=[('preprocessor',
                                        ColumnTransformer(transformers=[('num',
                                                                         Pipeline(steps=[('imputer',
                                                                                          SimpleImputer(strategy='median')),
                                                                                         ('scaler',
                                                                                          StandardScaler())]),
                                                                         <sklearn.compose._column_transformer.make_column_selector object at 0x7f8f36bdca90>),
                                                                        ('cat',
                                                                         OneHotEncoder(handle_unknown='ignore'),
                                                                         <sklearn.compose._column_transformer.make_column_selector object at 0x7f8f36bdcb80>)])),
                                       ('classifier', LogisticRegression())]),
             param_grid={'classifier__C': [0.1, 1.0, 10, 100],
                         'preprocessor__num__imputer__strategy': ['mean',
                                                                  'median']})
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.


Calling ‘fit’ triggers the cross-validated search for the best hyper-parameters combination:

grid_search.fit(X_train, y_train)

print("Best params:")
print(grid_search.best_params_)
Best params:
{'classifier__C': 0.1, 'preprocessor__num__imputer__strategy': 'mean'}

The internal cross-validation scores obtained by those parameters is:

print(f"Internal CV score: {grid_search.best_score_:.3f}")
Internal CV score: 0.784

We can also introspect the top grid search results as a pandas dataframe:

import pandas as pd

cv_results = pd.DataFrame(grid_search.cv_results_)
cv_results = cv_results.sort_values("mean_test_score", ascending=False)
cv_results[
    [
        "mean_test_score",
        "std_test_score",
        "param_preprocessor__num__imputer__strategy",
        "param_classifier__C",
    ]
].head(5)
mean_test_score std_test_score param_preprocessor__num__imputer__strategy param_classifier__C
0 0.784167 0.035824 mean 0.1
2 0.780366 0.032722 mean 1.0
1 0.780348 0.037245 median 0.1
4 0.779414 0.033105 mean 10
6 0.779414 0.033105 mean 100


The best hyper-parameters have be used to re-fit a final model on the full training set. We can evaluate that final model on held out test data that was not used for hyperparameter tuning.

print(
    (
        "best logistic regression from grid search: %.3f"
        % grid_search.score(X_test, y_test)
    )
)
best logistic regression from grid search: 0.794

Total running time of the script: ( 0 minutes 1.577 seconds)

Gallery generated by Sphinx-Gallery