westurner/sklearn-pandas

Pandas integration with sklearn

Sklearn-pandas

This module provides a bridge between Scikit-Learn's machine learning methods and pandas-style Data Frames.

In particular, it provides:

1. A way to map DataFrame columns to transformations, which are later recombined into features.
2. A compatibility shim for old scikit-learn versions to cross-validate a pipeline that takes a pandas DataFrame as input. This is only needed for scikit-learn<0.16.0 (see #11 for details). It is deprecated and will likely be dropped in skearn-pandas==2.0.
3. A couple of special transformers that work well with pandas inputs: CategoricalImputer and ``FunctionTransformer`.`

Installation

You can install sklearn-pandas with pip:

# pip install sklearn-pandas

Tests

The examples in this file double as basic sanity tests. To run them, use doctest, which is included with python:

# python -m doctest README.rst

Usage

Import

Import what you need from the sklearn_pandas package. The choices are:

• DataFrameMapper, a class for mapping pandas data frame columns to different sklearn transformations
• cross_val_score, similar to sklearn.cross_validation.cross_val_score but working on pandas DataFrames

For this demonstration, we will import both:

>>> from sklearn_pandas import DataFrameMapper, cross_val_score

For these examples, we'll also use pandas, numpy, and sklearn:

>>> import pandas as pd
>>> import numpy as np
>>> import sklearn.preprocessing, sklearn.decomposition, \
...     sklearn.linear_model, sklearn.pipeline, sklearn.metrics
>>> from sklearn.feature_extraction.text import CountVectorizer

Load some Data

Normally you'll read the data from a file, but for demonstration purposes we'll create a data frame from a Python dict:

>>> data = pd.DataFrame({'pet':      ['cat', 'dog', 'dog', 'fish', 'cat', 'dog', 'cat', 'fish'],
...                      'children': [4., 6, 3, 3, 2, 3, 5, 4],
...                      'salary':   [90., 24, 44, 27, 32, 59, 36, 27]})

Transformation Mapping

Map the Columns to Transformations

The mapper takes a list of tuples. The first element of each tuple is a column name from the pandas DataFrame, or a list containing one or multiple columns (we will see an example with multiple columns later). The second element is an object which will perform the transformation which will be applied to that column. The third one is optional and is a dictionary containing the transformation options, if applicable (see "custom column names for transformed features" below).

Let's see an example:

>>> mapper = DataFrameMapper([
...     ('pet', sklearn.preprocessing.LabelBinarizer()),
...     (['children'], sklearn.preprocessing.StandardScaler())
... ])

The difference between specifying the column selector as 'column' (as a simple string) and ['column'] (as a list with one element) is the shape of the array that is passed to the transformer. In the first case, a one dimensional array will be passed, while in the second case it will be a 2-dimensional array with one column, i.e. a column vector.

This behaviour mimics the same pattern as pandas' dataframes __getitem__ indexing:

>>> data['children'].shape
(8,)
>>> data[['children']].shape
(8, 1)

Be aware that some transformers expect a 1-dimensional input (the label-oriented ones) while some others, like OneHotEncoder or Imputer, expect 2-dimensional input, with the shape [n_samples, n_features].

Test the Transformation

We can use the fit_transform shortcut to both fit the model and see what transformed data looks like. In this and the other examples, output is rounded to two digits with np.round to account for rounding errors on different hardware:

>>> np.round(mapper.fit_transform(data.copy()), 2)
array([[ 1.  ,  0.  ,  0.  ,  0.21],
       [ 0.  ,  1.  ,  0.  ,  1.88],
       [ 0.  ,  1.  ,  0.  , -0.63],
       [ 0.  ,  0.  ,  1.  , -0.63],
       [ 1.  ,  0.  ,  0.  , -1.46],
       [ 0.  ,  1.  ,  0.  , -0.63],
       [ 1.  ,  0.  ,  0.  ,  1.04],
       [ 0.  ,  0.  ,  1.  ,  0.21]])

Note that the first three columns are the output of the LabelBinarizer (corresponding to cat, dog, and fish respectively) and the fourth column is the standardized value for the number of children. In general, the columns are ordered according to the order given when the DataFrameMapper is constructed.

Now that the transformation is trained, we confirm that it works on new data:

>>> sample = pd.DataFrame({'pet': ['cat'], 'children': [5.]})
>>> np.round(mapper.transform(sample), 2)
array([[1.  , 0.  , 0.  , 1.04]])

Output features names

In certain cases, like when studying the feature importances for some model, we want to be able to associate the original features to the ones generated by the dataframe mapper. We can do so by inspecting the automatically generated transformed_names_ attribute of the mapper after transformation:

>>> mapper.transformed_names_
['pet_cat', 'pet_dog', 'pet_fish', 'children']

Custom column names for transformed features

We can provide a custom name for the transformed features, to be used instead of the automatically generated one, by specifying it as the third argument of the feature definition:

>>> mapper_alias = DataFrameMapper([
...     (['children'], sklearn.preprocessing.StandardScaler(),
...      {'alias': 'children_scaled'})
... ])
>>> _ = mapper_alias.fit_transform(data.copy())
>>> mapper_alias.transformed_names_
['children_scaled']

Passing Series/DataFrames to the transformers

By default the transformers are passed a numpy array of the selected columns as input. This is because sklearn transformers are historically designed to work with numpy arrays, not with pandas dataframes, even though their basic indexing interfaces are similar.

However we can pass a dataframe/series to the transformers to handle custom cases initializing the dataframe mapper with input_df=True:

>>> from sklearn.base import TransformerMixin
>>> class DateEncoder(TransformerMixin):
...    def fit(self, X, y=None):
...        return self
...
...    def transform(self, X):
...        dt = X.dt
...        return pd.concat([dt.year, dt.month, dt.day], axis=1)
>>> dates_df = pd.DataFrame(
...     {'dates': pd.date_range('2015-10-30', '2015-11-02')})
>>> mapper_dates = DataFrameMapper([
...     ('dates', DateEncoder())
... ], input_df=True)
>>> mapper_dates.fit_transform(dates_df)
array([[2015,   10,   30],
       [2015,   10,   31],
       [2015,   11,    1],
       [2015,   11,    2]])

We can also specify this option per group of columns instead of for the whole mapper:

>>> mapper_dates = DataFrameMapper([
...     ('dates', DateEncoder(), {'input_df': True})
... ])
>>> mapper_dates.fit_transform(dates_df)
array([[2015,   10,   30],
       [2015,   10,   31],
       [2015,   11,    1],
       [2015,   11,    2]])

Outputting a dataframe

By default the output of the dataframe mapper is a numpy array. This is so because most sklearn estimators expect a numpy array as input. If however we want the output of the mapper to be a dataframe, we can do so using the parameter df_out when creating the mapper:

>>> mapper_df = DataFrameMapper([
...     ('pet', sklearn.preprocessing.LabelBinarizer()),
...     (['children'], sklearn.preprocessing.StandardScaler())
... ], df_out=True)
>>> np.round(mapper_df.fit_transform(data.copy()), 2)
   pet_cat  pet_dog  pet_fish  children
0        1        0         0      0.21
1        0        1         0      1.88
2        0        1         0     -0.63
3        0        0         1     -0.63
4        1        0         0     -1.46
5        0        1         0     -0.63
6        1        0         0      1.04
7        0        0         1      0.21

The names for the columns are the same ones present in the transformed_names_ attribute.

Note this does not work together with the default=True or sparse=True arguments to the mapper.

Transform Multiple Columns

Transformations may require multiple input columns. In these cases, the column names can be specified in a list:

>>> mapper2 = DataFrameMapper([
...     (['children', 'salary'], sklearn.decomposition.PCA(1))
... ])

Now running fit_transform will run PCA on the children and salary columns and return the first principal component:

>>> np.round(mapper2.fit_transform(data.copy()), 1)
array([[ 47.6],
       [-18.4],
       [  1.6],
       [-15.4],
       [-10.4],
       [ 16.6],
       [ -6.4],
       [-15.4]])

Multiple transformers for the same column

Multiple transformers can be applied to the same column specifying them in a list:

>>> mapper3 = DataFrameMapper([
...     (['age'], [sklearn.preprocessing.Imputer(),
...                sklearn.preprocessing.StandardScaler()])])
>>> data_3 = pd.DataFrame({'age': [1, np.nan, 3]})
>>> mapper3.fit_transform(data_3)
array([[-1.22474487],
       [ 0.        ],
       [ 1.22474487]])

Columns that don't need any transformation

Only columns that are listed in the DataFrameMapper are kept. To keep a column but don't apply any transformation to it, use None as transformer:

>>> mapper3 = DataFrameMapper([
...     ('pet', sklearn.preprocessing.LabelBinarizer()),
...     ('children', None)
... ])
>>> np.round(mapper3.fit_transform(data.copy()))
array([[1., 0., 0., 4.],
       [0., 1., 0., 6.],
       [0., 1., 0., 3.],
       [0., 0., 1., 3.],
       [1., 0., 0., 2.],
       [0., 1., 0., 3.],
       [1., 0., 0., 5.],
       [0., 0., 1., 4.]])

Applying a default transformer

A default transformer can be applied to columns not explicitly selected passing it as the default argument to the mapper:

>>> mapper4 = DataFrameMapper([
...     ('pet', sklearn.preprocessing.LabelBinarizer()),
...     ('children', None)
... ], default=sklearn.preprocessing.StandardScaler())
>>> np.round(mapper4.fit_transform(data.copy()), 1)
array([[ 1. ,  0. ,  0. ,  4. ,  2.3],
       [ 0. ,  1. ,  0. ,  6. , -0.9],
       [ 0. ,  1. ,  0. ,  3. ,  0.1],
       [ 0. ,  0. ,  1. ,  3. , -0.7],
       [ 1. ,  0. ,  0. ,  2. , -0.5],
       [ 0. ,  1. ,  0. ,  3. ,  0.8],
       [ 1. ,  0. ,  0. ,  5. , -0.3],
       [ 0. ,  0. ,  1. ,  4. , -0.7]])

Using default=False (the default) drops unselected columns. Using default=None pass the unselected columns unchanged.

Same transformer for the multiple columns

Sometimes it is required to apply the same transformation to several dataframe columns. To simplify this process, the package provides gen_features function which accepts a list of columns and feature transformer class (or list of classes), and generates a feature definition, acceptable by DataFrameMapper.

For example, consider a dataset with three categorical columns, 'col1', 'col2', and 'col3', To binarize each of them, one could pass column names and LabelBinarizer transformer class into generator, and then use returned definition as features argument for DataFrameMapper:

>>> from sklearn_pandas import gen_features
>>> feature_def = gen_features(
...     columns=['col1', 'col2', 'col3'],
...     classes=[sklearn.preprocessing.LabelEncoder]
... )
>>> feature_def
[('col1', [LabelEncoder()]), ('col2', [LabelEncoder()]), ('col3', [LabelEncoder()])]
>>> mapper5 = DataFrameMapper(feature_def)
>>> data5 = pd.DataFrame({
...     'col1': ['yes', 'no', 'yes'],
...     'col2': [True, False, False],
...     'col3': ['one', 'two', 'three']
... })
>>> mapper5.fit_transform(data5)
array([[1, 1, 0],
       [0, 0, 2],
       [1, 0, 1]])

If it is required to override some of transformer parameters, then a dict with 'class' key and transformer parameters should be provided. For example, consider a dataset with missing values. Then the following code could be used to override default imputing strategy:

>>> feature_def = gen_features(
...     columns=[['col1'], ['col2'], ['col3']],
...     classes=[{'class': sklearn.preprocessing.Imputer, 'strategy': 'most_frequent'}]
... )
>>> mapper6 = DataFrameMapper(feature_def)
>>> data6 = pd.DataFrame({
...     'col1': [None, 1, 1, 2, 3],
...     'col2': [True, False, None, None, True],
...     'col3': [0, 0, 0, None, None]
... })
>>> mapper6.fit_transform(data6)
array([[1., 1., 0.],
       [1., 0., 0.],
       [1., 1., 0.],
       [2., 1., 0.],
       [3., 1., 0.]])

Feature selection and other supervised transformations

DataFrameMapper supports transformers that require both X and y arguments. An example of this is feature selection. Treating the 'pet' column as the target, we will select the column that best predicts it.

>>> from sklearn.feature_selection import SelectKBest, chi2
>>> mapper_fs = DataFrameMapper([(['children','salary'], SelectKBest(chi2, k=1))])
>>> mapper_fs.fit_transform(data[['children','salary']], data['pet'])
array([[90.],
       [24.],
       [44.],
       [27.],
       [32.],
       [59.],
       [36.],
       [27.]])

Working with sparse features

A DataFrameMapper will return a dense feature array by default. Setting sparse=True in the mapper will return a sparse array whenever any of the extracted features is sparse. Example:

>>> mapper5 = DataFrameMapper([
...     ('pet', CountVectorizer()),
... ], sparse=True)
>>> type(mapper5.fit_transform(data))
<class 'scipy.sparse.csr.csr_matrix'>

The stacking of the sparse features is done without ever densifying them.

Cross-Validation

Now that we can combine features from pandas DataFrames, we may want to use cross-validation to see whether our model works. scikit-learn<0.16.0 provided features for cross-validation, but they expect numpy data structures and won't work with DataFrameMapper.

To get around this, sklearn-pandas provides a wrapper on sklearn's cross_val_score function which passes a pandas DataFrame to the estimator rather than a numpy array:

>>> pipe = sklearn.pipeline.Pipeline([
...     ('featurize', mapper),
...     ('lm', sklearn.linear_model.LinearRegression())])
>>> np.round(cross_val_score(pipe, X=data.copy(), y=data.salary, scoring='r2'), 2)
array([ -1.09,  -5.3 , -15.38])

Sklearn-pandas' cross_val_score function provides exactly the same interface as sklearn's function of the same name.

CategoricalImputer

Since the scikit-learn Imputer transformer currently only works with numbers, sklearn-pandas provides an equivalent helper transformer that works with strings, substituting null values with the most frequent value in that column. Alternatively, you can specify a fixed value to use.

Example: imputing with the mode:

>>> from sklearn_pandas import CategoricalImputer
>>> data = np.array(['a', 'b', 'b', np.nan], dtype=object)
>>> imputer = CategoricalImputer()
>>> imputer.fit_transform(data)
array(['a', 'b', 'b', 'b'], dtype=object)

Example: imputing with a fixed value:

>>> from sklearn_pandas import CategoricalImputer
>>> data = np.array(['a', 'b', 'b', np.nan], dtype=object)
>>> imputer = CategoricalImputer(strategy='constant', fill_value='a')
>>> imputer.fit_transform(data)
array(['a', 'b', 'b', 'a'], dtype=object)

FunctionTransformer

Often one wants to apply simple transformations to data such as np.log. FunctionTransformer is a simple wrapper that takes any function and applies vectorization so that it can be used as a transformer.

Example:

>>> from sklearn_pandas import FunctionTransformer
>>> array = np.array([10, 100])
>>> transformer = FunctionTransformer(np.log10)

>>> transformer.fit_transform(array)
array([1., 2.])

Changelog

1.8.0 (2018-12-01)

• Add FunctionTransformer class (#117).
• Fix column names derivation for dataframes with multi-index or non-string columns (#166).
• Change behaviour of DataFrameMapper's fit_transform method to invoke each underlying transformers' native fit_transform if implemented. (#150)

1.7.0 (2018-08-15)

• Fix issues with unicode names in get_names (#160).
• Update to build using numpy==1.14 and python==3.6 (#154).
• Add strategy and fill_value parameters to CategoricalImputer to allow imputing with values other than the mode (#144), (#161).
• Preserve input data types when no transform is supplied (#138).

1.6.0 (2017-10-28)

• Add column name to exception during fit/transform (#110).
• Add gen_feature helper function to help generating the same transformation for multiple columns (#126).

1.5.0 (2017-06-24)

• Allow inputting a dataframe/series per group of columns.
• Get feature names also from estimator.get_feature_names() if present.
• Attempt to derive feature names from individual transformers when applying a list of transformers.
• Do not mutate features in __init__ to be compatible with sklearn>=0.20 (#76).

1.4.0 (2017-05-13)

• Allow specifying a custom name (alias) for transformed columns (#83).
• Capture output columns generated names in transformed_names_ attribute (#78).
• Add CategoricalImputer that replaces null-like values with the mode for string-like columns.
• Add input_df init argument to allow inputting a dataframe/series to the transformers instead of a numpy array (#60).

1.3.0 (2017-01-21)

• Make the mapper return dataframes when df_out=True (#70, #74).
• Update imports to avoid deprecation warnings in sklearn 0.18 (#68).

1.2.0 (2016-10-02)

• Deprecate custom cross-validation shim classes.
• Require scikit-learn>=0.15.0. Resolves #49.
• Allow applying a default transformer to columns not selected explicitly in the mapper. Resolves #55.
• Allow specifying an optional y argument during transform for supervised transformations. Resolves #58.

1.1.0 (2015-12-06)

• Delete obsolete PassThroughTransformer. If no transformation is desired for a given column, use None as transformer.
• Factor out code in several modules, to avoid having everything in __init__.py.
• Use custom TransformerPipeline class to allow transformation steps accepting only a X argument. Fixes #46.
• Add compatibility shim for unpickling mappers with list of transformers created before 1.0.0. Fixes #45.

1.0.0 (2015-11-28)

• Change version numbering scheme to SemVer.
• Use sklearn.pipeline.Pipeline instead of copying its code. Resolves #43.
• Raise KeyError when selecting unexistent columns in the dataframe. Fixes #30.
• Return sparse feature array if any of the features is sparse and sparse argument is True. Defaults to False to avoid potential breaking of existing code. Resolves #34.
• Return model and prediction in custom CV classes. Fixes #27.

0.0.12 (2015-11-07)

• Allow specifying a list of transformers to use sequentially on the same column.

Credits

The code for DataFrameMapper is based on code originally written by Ben Hamner.

Other contributors:

• Ariel Rossanigo (@arielrossanigo)
• Arnau Gil Amat (@arnau126)
• Assaf Ben-David (@AssafBenDavid)
• Brendan Herger (@bjherger)
• Cal Paterson (@calpaterson)
• @defvorfu
• Gustavo Sena Mafra (@gsmafra)
• Israel Saeta Pérez (@dukebody)
• Jeremy Howard (@jph00)
• Jimmy Wan (@jimmywan)
• Kristof Van Engeland (@kristofve91)
• Olivier Grisel (@ogrisel)
• Paul Butler (@paulgb)
• Richard Miller (@rwjmiller)
• Ritesh Agrawal (@ragrawal)
• @SandroCasagrande
• Timothy Sweetser (@hacktuarial)
• Vitaley Zaretskey (@vzaretsk)
• Zac Stewart (@zacstewart)