ColumnTransformer(s) in various parts of a pipeline do not play well

Question

I am using sklearn and mlxtend.regressor.StackingRegressor to build a stacked regression model. For example, say I want the following small pipeline:

A Stacking Regressor with two regressors:
- A pipeline which:
  - Performs data imputation
  - 1-hot encodes categorical features
  - Performs linear regression
- A pipeline which:
  - Performs data imputation
  - Performs regression using a Decision Tree

Unfortunately this is not possible, because StackingRegressor doesn't accept NaN in its input data. This is even if its regressors know how to handle NaN, as it would be in my case where the regressors are actually pipelines which perform data imputation.

However, this is not a problem: I can just move data imputation outside the stacked regressor. Now my pipeline looks like this:

Perform data imputation
Apply a Stacking Regressor with two regressors:
- A pipeline which:
  - 1-hot encodes categorical features
  - Standardises numerical features
  - Performs linear regression
- An sklearn.tree.DecisionTreeRegressor.

One might try to implement it as follows (the entire minimal working example in this gist, with comments):

sr_linear = Pipeline(steps=[
    ('preprocessing', ColumnTransformer(transformers=[
        ('categorical',
             make_pipeline(OneHotEncoder(), StandardScaler()),
             make_column_selector(dtype_include='category')),
        ('numerical',
             StandardScaler(),
             make_column_selector(dtype_include=np.number))
    ])),
    ('model', LinearRegression())
])

sr_tree = DecisionTreeRegressor()

ct_imputation = ColumnTransformer(transformers=[
    ('categorical',
        SimpleImputer(strategy='constant', fill_value='None'),
        make_column_selector(dtype_include='category')),
    ('numerical',
        SimpleImputer(strategy='median'),
        make_column_selector(dtype_include=np.number))
])

stacked_regressor = Pipeline(steps=[
    ('imputation', ct_imputation),
    ('back_to_pandas', FunctionTransformer(
        func=lambda values: pd.DataFrame(values, columns=ct_imputation.get_feature_names_out())
    )),
    ('model', StackingRegressor(
        regressors=[sr_linear, sr_tree],
        meta_regressor=DecisionTreeRegressor(),
        use_features_in_secondary=True
    ))
])

Note that the "outer" ColumnTransformer (in stacked_regressor) returns a numpy matrix. But the "inner" ColumnTransformer (in sr_linear) expects a pandas.DataFrame, so I had to convert the matrix back to a data frame using step back_to_pandas. (To use get_feature_names_out I had to use the nightly version of sklearn, because the current stable 1.0.2 version does not support it yet. Fortunately it can be installed with one simple command.)

The above code fails when calling stacked_regressor.fit(), with the following error (the entire stacktrace is again in the gist):

ValueError: make_column_selector can only be applied to pandas dataframes

However, because I added the back_to_pandas step to my outer pipeline, the inner pipelines should be getting a pandas data frame! In fact, if I only fit_transform() my ct_imputation object, I clearly obtain a pandas data frame. I cannot understand where and when exactly the data which gets passed around ceases to be a data frame. Why is my code failing?

amiola · Answer 1 · 2022-02-18T21:31:49.927

Imo the issue has to be ascribed to StackingRegressor. Actually, I am not an expert on its usage and still I have not explored its source code, but I've found this sklearn issue - #16473 which seems implying that << the concatenation [of regressors and meta_regressors] does not preserve dataframe >> (though this is referred to sklearn StackingRegressor instance, rather than on mlxtend one).

Indeed, have a look at what happens once you replace it with your sr_linear pipeline:

from sklearn.datasets import fetch_openml
from sklearn.pipeline import Pipeline, make_pipeline
from sklearn.compose import ColumnTransformer, make_column_selector
from sklearn.impute import SimpleImputer
from sklearn.preprocessing import StandardScaler, OneHotEncoder, FunctionTransformer
from sklearn.linear_model import LinearRegression
from sklearn.tree import DecisionTreeRegressor
from sklearn.model_selection import train_test_split

from mlxtend.regressor import StackingRegressor

import numpy as np
import pandas as pd

# We use the Ames house prices dataset for this example
d = fetch_openml('house_prices', as_frame=True).frame

# Small data preprocessing:
for column in d.columns:
    if d[column].dtype == object or column == 'MSSubClass':
        d[column] = pd.Categorical(d[column])
    
d.drop(columns='Id', inplace=True)

# Prepare the data for training
label = 'SalePrice'
features = [col for col in d.columns if col != label]
X, y = d[features], d[label]

# Train the stacked regressor
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, shuffle=True)

sr_linear = Pipeline(steps=[
('preprocessing', ColumnTransformer(transformers=[
    ('categorical',
         make_pipeline(OneHotEncoder(), StandardScaler(with_mean=False)),
         make_column_selector(dtype_include='category')),
     ('numerical',
         StandardScaler(),
         make_column_selector(dtype_include=np.number))
    ])),
    ('model', LinearRegression())
])

ct_imputation = ColumnTransformer(transformers=[
    ('categorical',
        SimpleImputer(missing_values=np.nan, strategy='constant', fill_value='None'),
        make_column_selector(dtype_include='category')),
    ('numerical',
        SimpleImputer(strategy='median'),
        make_column_selector(dtype_include=np.number))
])

stacked_regressor = Pipeline(steps=[
    ('imputation', ct_imputation),
    ('back_to_pandas', FunctionTransformer(
        func=lambda values: pd.DataFrame(values, columns=ct_imputation.get_feature_names_out()).astype(types)
    )),
    ('mdl', sr_linear)
])

stacked_regressor.fit(X_train, y_train)

Observe that I had to slightly modify the 'back_to_pandas' step because for some reason pd.DataFrame was changing the dtypes of the columns to 'object' only (from 'category' and 'float64'), therefore clashing with the imputation performed in sr_linear. For this reason, I've applied .astype(types) to the pd.DataFrame constructor, where types is defined as follows (based on the implementation of .get_feature_names_out() method of the SimpleImputer from the dev version of sklearn):

types = {} 
for col in d.columns[:-1]: 
    if d[col].dtype == 'category':
        types['categorical__' + col] = str(d[col].dtype)
    else:
        types['numerical__' + col] = str(d[col].dtype)

Good catch about the column types. I had already investigated `mlxtend` source and I had not found anything suspicious. I see that removing the stacked regressor makes things work. I also replaced mlxtend's object with sklearn's one and, as you say, I "lose" the dataframe and this confuses the models: `ValueError: could not convert string to float: 'RL'`. The only solution is possibly to write my own StackingRegressor object. — Alberto Santini, Feb 19 '22 at 19:22
Ah, wait! That error is just because sklearn's `DecisionTreeRegressor` does not support categorical features. After one-hot-encoding it actually works. I will post an answer with the solution. — Alberto Santini, Feb 19 '22 at 19:34

score 1 · Accepted Answer · answered Feb 19 '22 at 19:40

The correct thing to do was:

Move from mlxtend's to sklearn's StackingRegressor. I believe the former was creater when sklearn still didn't have a stacking regressor. Now there is no need to use more 'obscure' solutions. sklearn's stacking regressor works pretty well.
Move the 1-hot-encoding step to the outer pipeline, because (surprisingly!) sklearn's DecisionTreeRegressor cannot handle categorical data among the features.

A working version of the code is given below:

from sklearn.datasets import fetch_openml
from sklearn.pipeline import Pipeline, make_pipeline
from sklearn.compose import ColumnTransformer, make_column_selector
from sklearn.impute import SimpleImputer
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.linear_model import LinearRegression
from sklearn.tree import DecisionTreeRegressor
from sklearn.model_selection import train_test_split
from sklearn.ensemble import StackingRegressor

import numpy as np
import pandas as pd

def set_correct_categories(df: pd.DataFrame) -> pd.DataFrame:
    for column in df.columns:
        if df[column].dtype == object or 'MSSubClass' in column:
            df[column] = pd.Categorical(df[column])

    return df

d = fetch_openml('house_prices', as_frame=True).frame
d = set_correct_categories(d).drop(columns='Id')

sr_linear = Pipeline(steps=[
    ('preprocessing', StandardScaler()),
    ('model', LinearRegression())
])

ct_preprocessing = ColumnTransformer(transformers=[
    ('categorical',
        make_pipeline(
            SimpleImputer(strategy='constant', fill_value='None'),
            OneHotEncoder(sparse=False, handle_unknown='ignore')
        ),
        make_column_selector(dtype_include='category')),
    ('numerical',
        SimpleImputer(strategy='median'),
        make_column_selector(dtype_include=np.number))
])

stacking_regressor = Pipeline(steps=[
    ('preprocessing', ct_preprocessing),
    ('model', StackingRegressor(
        estimators=[('linear_regression', sr_linear), ('regression_tree', DecisionTreeRegressor())],
        final_estimator=DecisionTreeRegressor(),
        passthrough=True
    ))
])

label = 'SalePrice'
features = [col for col in d.columns if col != label]
X, y = d[features], d[label]
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, shuffle=True)

stacking_regressor.fit(X_train, y_train)

Thanks to user amiola for his answer putting me on the right track.

ColumnTransformer(s) in various parts of a pipeline do not play well

2 Answers2