Description

I am using ColumnTransformer to prepare (impute etc.) heterogenous data. I use a DataFrame to have more control on the different (types of) columns by their name.

I had some really cryptic problems when downstream transformers complained of data of the wrong type, while the ColumnTransformer should have divided them up properly.

I found out that ColumnTransformer silently passes the wrong columns along as remainder when

• specifying columns by name,
• using the remainder option, and using
• DataFrames where column ordering can differ between fit and transform

In this case, the wrong columns are passed on to the downstream transformers, as the example demonstrates:

Steps/Code to Reproduce

from sklearn.compose import make_column_transformer
from sklearn.preprocessing import FunctionTransformer
import pandas as pd

def msg(msg):
  def print_cols(X, y=None):
    print(msg, list(X.columns))
    return X
  return print_cols

ct = make_column_transformer(
  (FunctionTransformer(msg('col a'), validate=False), ['a']),
  remainder=FunctionTransformer(msg('remainder'), validate=False)
)

fit_df = pd.DataFrame({
  'a': [2,3], 
  'b': [4,5]})

ct.fit(fit_df)

# prints:
# cols a ['a']
# remainder ['b']

transform_df = pd.DataFrame({
  'b': [4,5],  # note that column ordering
  'a': [2,3]}) # is the only difference to fit_df

ct.transform(transform_df)

# prints:
# col a ['a']
# remainder ['a'] <-- Should be ['b']

Expected Results

When using ColumnTransformer with a DataFrame and specifying columns by name, remainder should reference the same columns when fitting and when transforming (['b'] in above example), regardless of the column positions in the data during fitting and transforming.

Actual Results

remainder appears to, during fitting, remember remaining named DataFrame columns by their numeric index (not by their names), which (silently) leads to the wrong columns being handled downstream if the transformed DataFrame's column ordering differs from that of the fitted DataFrame.

Position in module where the remainder index is determined:

My current workaround is to not use the remainder option but specify all columns by name explicitly.

Versions

System:
python: 3.7.3 (default, Mar 30 2019, 03:44:34) [Clang 9.1.0 (clang-902.0.39.2)]
executable: /Users/asschude/.local/share/virtualenvs/launchpad-mmWds3ry/bin/python
machine: Darwin-17.7.0-x86_64-i386-64bit

BLAS:
macros: NO_ATLAS_INFO=3, HAVE_CBLAS=None
lib_dirs:
cblas_libs: cblas

Python deps:
pip: 19.1.1
setuptools: 41.0.1
sklearn: 0.21.2
numpy: 1.16.4
scipy: 1.3.0
Cython: None
pandas: 0.24.2