Skip to main content

Pandas Adapters For Scikit-Learn

Project description

Ami Tavory, Shahar Azulay, Tali Raveh-Sadka

This library aims for two (somewhat independent) goals:

(You might also want to check out the excellent pandas-sklearn which has the same aims, but takes a very different approach.)

The full documentation at rtd_ibex defines these matters in detail, but the library has an extremely-small interface.


The following short example shows the main points of the library. It is an adaptation of the scikit-learn example Concatenating multiple feature extraction methods. In this example, we build a classifier for the iris dataset using a combination of PCA, univariate feature selection, and a support vecor machine classifier.

We first load the Iris dataset into a pandas DataFrame.

>>> import numpy as np
>>> from sklearn import datasets
>>> import pandas as pd
>>> iris = datasets.load_iris()
>>> features, targets, iris = iris['feature_names'], iris['target_names'], pd.DataFrame(
...     np.c_[iris['data'], iris['target']],
...     columns=iris['feature_names']+['class'])
>>> iris['class'] = iris['class'].map(pd.Series(targets))
>>> iris.head()
   sepal length (cm)  sepal width (cm)  petal length (cm)  petal width (cm)  \
0                5.1               3.5                1.4               0.2
1                4.9               3.0                1.4               0.2
2                4.7               3.2                1.3               0.2
3                4.6               3.1                1.5               0.2
4                5.0               3.6                1.4               0.2
0  setosa
1  setosa
2  setosa
3  setosa
4  setosa

Now, we import the relevant steps. Note that, in this example, we import them from ibex.sklearn rather than sklearn.

>>> from ibex.sklearn.svm import SVC as PdSVC
>>> from ibex.sklearn.feature_selection import SelectKBest as PdSelectKBest
>>> from ibex.sklearn.decomposition import PCA as PdPCA

(Of course, it’s possible to import steps from sklearn as well, and use them alongside and together with the steps of ibex.sklearn.)

Finally, we construct a pipeline that, given a DataFrame of features:

  • horizontally concatenates a 2-component PCA DataFrame, and the best-feature DataFrame, to a resulting DataFrame

  • then, passes the result to a support-vector machine classifier outputting a pandas series:

    >>> clf = PdPCA(n_components=2) + PdSelectKBest(k=1) | PdSVC(kernel="linear")

clf is now a pandas-ware classifier, but otherwise can be used pretty much like all sklearn estimator. For example,

>>> param_grid = dict(
...     featureunion__pca__n_components=[1, 2, 3],
...     featureunion__selectkbest__k=[1, 2],
...     svc__C=[0.1, 1, 10])
>>> try:
...     from ibex.sklearn.model_selection import GridSearchCV as PdGridSearchCV
... except: # Accomodate older versions of sklearn
...     from ibex.sklearn.grid_search import GridSearchCV as PdGridSearchCV
>>> PdGridSearchCV(clf, param_grid=param_grid).fit(iris[features], iris['class']) # doctest: +SKIP

So what does this add to the original version?

  1. The estimators perform verification and processing on the inputs and outputs. They verify column names following calls to fit, and index results according to those of the inputs. This helps catch bugs.

  2. The results are much more interpretable:

    >>> svc = PdSVC(kernel="linear", probability=True)

    Find the coefficients of the boundaries between the different classes:

    >>>[features], iris['class']).coef_
                sepal length (cm)  sepal width (cm)  petal length (cm)  \
    setosa              -0.046259          0.521183          -1.003045
    versicolor          -0.007223          0.178941          -0.538365
    virginica            0.595498          0.973900          -2.031000
                petal width (cm)
    setosa             -0.464130
    versicolor         -0.292393
    virginica          -2.006303

    Predict belonging to classes:

    >>>[features], iris['class']).predict_proba(iris[features])
        setosa  versicolor  virginica
    0    0.97...    0.01...   0.00...

    Find the coefficients of the boundaries between the different classes in a pipeline:

    >>> clf = PdPCA(n_components=2) + PdSelectKBest(k=1) | svc
    >>> clf =[features], iris['class'])
    >>> svc.coef_
                    pca                 selectkbest
                comp_0    comp_1 petal length (cm)
    setosa     -0.757016  ...0.376680         -0.575197
    versicolor -0.351218  ...0.141699         -0.317562
    virginica  -1.529320  ...1.472771         -1.509391
  3. It allows writinfitg Pandas-munging estimators (see also Multiple-Row Features In The Movielens Dataset).

  4. Using DataFrame metadata, it allows writing more complex meta-learning algorithms, such as stacking and nested labeled and stratified cross validation.

  5. The pipeline syntax is succinct and clear (see Motivation For Shorter Combinations).

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

ibex-0.1.3.tar.gz (24.5 kB view hashes)

Uploaded source

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Huawei Huawei PSF Sponsor Microsoft Microsoft PSF Sponsor NVIDIA NVIDIA PSF Sponsor Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page