Feature-engine

Common errors

• TypeError: Some of the variables are not categorical. Please cast them as object or category before calling this transformer

  cause A categorical encoding transformer was applied to variables that are not of pandas 'object' or 'category' dtype, but rather numerical.
  fix Convert the target numerical column(s) to 'object' or 'category' dtype before fitting the encoder, or set `ignore_format=True` in the encoder's constructor if you deliberately want to encode numerical columns (e.g., `df['col'] = df['col'].astype('object')`).

• KeyError: "['some_category'] not in index"

  cause This often occurs when trying to directly access or manipulate categories that are present in the test set but were not seen in the training set by a fitted encoder or discretizer.
  fix Use `feature_engine.preprocessing.MatchCategories()` at the preprocessing stage to ensure consistent categories across train and test sets. Alternatively, review your chosen encoder's parameters for handling unknown categories (e.g., `RareLabelEncoder(tol=...)`, `OneHotEncoder(handle_unknown='ignore')`).

Warnings

• breaking Module paths and some class names were renamed in v1.0.0 to better reflect their functionality and align with Scikit-learn conventions. For example, `feature_engine.imputers` became `feature_engine.imputation`.
  Fix: Update import statements to the new module paths. Refer to the official documentation or the v1.0.0 release notes for a complete list of changes. E.g., `from feature_engine.imputation import MeanMedianImputer`.
• gotcha Feature-engine's categorical encoders (e.g., `MeanEncoder`, `OneHotEncoder`) by default expect variables to be of pandas `object` or `category` dtype. Providing numerical variables without explicit handling will raise a `TypeError`.
  Fix: Ensure target variables are cast to `object` or `category` dtype before applying the encoder, or set the `ignore_format=True` parameter in the transformer if you intend to encode numerical variables (use with caution). Example: `df['numerical_col'] = df['numerical_col'].astype('object')`.
• gotcha When dealing with categorical features, it's common for the test set to contain categories not present in the training set ('unseen categories'). This can lead to errors during transformation, especially with certain encoding schemes.
  Fix: Use `feature-engine`'s `MatchCategories` transformer (introduced in v1.5.0) as part of your pipeline to align categories between training and test sets. Alternatively, ensure your chosen encoder has a strategy for handling unseen categories (e.g., `handle_unknown='ignore'` or a custom mapping). For `OneHotEncoder` specifically, a bug fix in v1.1.2 addressed how it handles binary variables with `drop_last_binary=True`.
• breaking In v1.1.0, most transformers gained a new attribute `variables_` which contains the names of the variables that were actually modified by the transformer. While the old `variables` attribute is generally retained, `variables_` should be preferred for consistency and accuracy.
  Fix: If your code inspects the list of variables transformed, switch to using `transformer.variables_` instead of `transformer.variables` for robust behavior, especially in complex pipelines where `variables` might refer to initial input and `variables_` to final processed ones.

Install

• pip install feature-engine PyPI
• conda install -c conda-forge feature_engine Anaconda/Conda-forge

Imports

• MeanMedianImputer
  wrong:

  from feature_engine.imputers import MeanMedianImputer

  correct:

  from feature_engine.imputation import MeanMedianImputer

  Module paths were renamed in v1.0.0. The correct path is now `feature_engine.imputation`.
• OneHotEncoder
  wrong:

  from feature_engine.categorical_encoders import OneHotEncoder

  correct:

  from feature_engine.encoding import OneHotEncoder

  Module paths were renamed in v1.0.0. The correct path is now `feature_engine.encoding`.
• DropCorrelatedFeatures
  wrong:

  from feature_engine.variable_selection import DropCorrelatedFeatures

  correct:

  from feature_engine.selection import DropCorrelatedFeatures

  Module paths were renamed in v1.0.0. The correct path is now `feature_engine.selection`.

Quickstart

This quickstart demonstrates how to use `feature-engine`'s `MeanMedianImputer` to handle missing data. It loads a dataset, splits it into training and testing sets, fits the imputer on the training data, and then transforms both sets. This follows the standard Scikit-learn `fit()` and `transform()` pattern, ensuring proper parameter learning from training data.

import pandas as pd
from sklearn.datasets import fetch_openml
from sklearn.model_selection import train_test_split
from feature_engine.imputation import MeanMedianImputer

# Load dataset
X, y = fetch_openml(name="house_prices", version=1, as_frame=True, return_X_y=True)

# Separate into train and test sets
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.3, random_state=0
)

# Initialize a MeanMedianImputer for specified numerical variables
# It will impute 'median' for LotFrontage and MasVnrArea
median_imputer = MeanMedianImputer(
    imputation_method='median',
    variables=['LotFrontage', 'MasVnrArea']
)

# Fit the imputer on the training data
median_imputer.fit(X_train)

# Transform both training and test data
X_train_imputed = median_imputer.transform(X_train)
X_test_imputed = median_imputer.transform(X_test)

print("Missing values in 'LotFrontage' before imputation (train):", X_train['LotFrontage'].isnull().sum())
print("Missing values in 'LotFrontage' after imputation (train):", X_train_imputed['LotFrontage'].isnull().sum())
print("Missing values in 'MasVnrArea' before imputation (test):", X_test['MasVnrArea'].isnull().sum())
print("Missing values in 'MasVnrArea' after imputation (test):", X_test_imputed['MasVnrArea'].isnull().sum())