This module provides a bridge between Scikit-Learn's machine learning methods and pandas-style Data Frames.
In particular, it provides:
- a way to map DataFrame columns to transformations, which are later recombined into features
- a way to cross-validate a pipeline that takes a pandas DataFrame as input.
You can install sklearn-pandas with pip:
# pip install sklearn-pandas
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
Import what you need from the sklearn_pandas package. The choices are:
DataFrameMapper, a class for mapping pandas data frame columns to different sklearn transformationscross_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
Normally you'll read the data from a file, but for demonstration purposes I'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]})
The mapper takes a list of pairs. The first 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 is an object which will perform the transformation which will be applied to that column:
>>> 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 with 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].
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]])
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 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]])
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.]])
Now that we can combine features from pandas DataFrames, we may want to use cross-validation to see whether our model works. Scikit-learn provides 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, data.copy(), data.salary, '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.
The code for DataFrameMapper is based on code originally written by Ben Hamner.