Show Sidebar Hide Sidebar

Feature Union with Heterogeneous Data Sources in Scikit-learn

Note: this page is part of the documentation for version 3 of, which is not the most recent version.
See our Version 4 Migration Guide for information about how to upgrade.

Datasets can often contain components of that require different feature extraction and processing pipelines. This scenario might occur when:

  • Your dataset consists of heterogeneous data types (e.g. raster images and text captions)
  • Your dataset is stored in a Pandas DataFrame and different columns require different processing pipelines.

This example demonstrates how to use sklearn.feature_extraction. FeatureUnion on a dataset containing different types of features. We use the 20-newsgroups dataset and compute standard bag-of-words features for the subject line and body in separate pipelines as well as ad hoc features on the body. We combine them (with weights) using a FeatureUnion and finally train a classifier on the combined set of features.

The choice of features is not particularly helpful, but serves to illustrate the technique.

New to Plotly?

Plotly's Python library is free and open source! Get started by downloading the client and reading the primer.
You can set up Plotly to work in online or offline mode, or in jupyter notebooks.
We also have a quick-reference cheatsheet (new!) to help you get started!


In [1]:
from __future__ import print_function
import numpy as np

from sklearn.base import BaseEstimator, TransformerMixin
from sklearn.datasets import fetch_20newsgroups
from sklearn.datasets.twenty_newsgroups import strip_newsgroup_footer
from sklearn.datasets.twenty_newsgroups import strip_newsgroup_quoting

from sklearn.decomposition import TruncatedSVD
from sklearn.feature_extraction import DictVectorizer
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics import classification_report

from sklearn.pipeline import FeatureUnion
from sklearn.pipeline import Pipeline
from sklearn.svm import SVC


In [ ]:
class ItemSelector(BaseEstimator, TransformerMixin):
    """For data grouped by feature, select subset of data at a provided key.

    The data is expected to be stored in a 2D data structure, where the first
    index is over features and the second is over samples.  i.e.

    >> len(data[key]) == n_samples

    Please note that this is the opposite convention to scikit-learn feature
    matrixes (where the first index corresponds to sample).

    ItemSelector only requires that the collection implement getitem
    (data[key]).  Examples include: a dict of lists, 2D numpy array, Pandas
    DataFrame, numpy record array, etc.

    >> data = {'a': [1, 5, 2, 5, 2, 8],
               'b': [9, 4, 1, 4, 1, 3]}
    >> ds = ItemSelector(key='a')
    >> data['a'] == ds.transform(data)

    ItemSelector is not designed to handle data grouped by sample.  (e.g. a
    list of dicts).  If your data is structured this way, consider a
    transformer along the lines of `sklearn.feature_extraction.DictVectorizer`.

    key : hashable, required
        The key corresponding to the desired value in a mappable.
    def __init__(self, key):
        self.key = key

    def fit(self, x, y=None):
        return self

    def transform(self, data_dict):
        return data_dict[self.key]

class TextStats(BaseEstimator, TransformerMixin):
    """Extract features from each document for DictVectorizer"""

    def fit(self, x, y=None):
        return self

    def transform(self, posts):
        return [{'length': len(text),
                 'num_sentences': text.count('.')}
                for text in posts]

class SubjectBodyExtractor(BaseEstimator, TransformerMixin):
    """Extract the subject & body from a usenet post in a single pass.

    Takes a sequence of strings and produces a dict of sequences.  Keys are
    `subject` and `body`.
    def fit(self, x, y=None):
        return self

    def transform(self, posts):
        features = np.recarray(shape=(len(posts),),
                               dtype=[('subject', object), ('body', object)])
        for i, text in enumerate(posts):
            headers, _, bod = text.partition('\n\n')
            bod = strip_newsgroup_footer(bod)
            bod = strip_newsgroup_quoting(bod)
            features['body'][i] = bod

            prefix = 'Subject:'
            sub = ''
            for line in headers.split('\n'):
                if line.startswith(prefix):
                    sub = line[len(prefix):]
            features['subject'][i] = sub

        return features

pipeline = Pipeline([
    # Extract the subject & body
    ('subjectbody', SubjectBodyExtractor()),

    # Use FeatureUnion to combine the features from subject and body
    ('union', FeatureUnion(

            # Pipeline for pulling features from the post's subject line
            ('subject', Pipeline([
                ('selector', ItemSelector(key='subject')),
                ('tfidf', TfidfVectorizer(min_df=50)),

            # Pipeline for standard bag-of-words model for body
            ('body_bow', Pipeline([
                ('selector', ItemSelector(key='body')),
                ('tfidf', TfidfVectorizer()),
                ('best', TruncatedSVD(n_components=50)),

            # Pipeline for pulling ad hoc features from post's body
            ('body_stats', Pipeline([
                ('selector', ItemSelector(key='body')),
                ('stats', TextStats()),  # returns a list of dicts
                ('vect', DictVectorizer()),  # list of dicts -> feature matrix


        # weight components in FeatureUnion
            'subject': 0.8,
            'body_bow': 0.5,
            'body_stats': 1.0,

    # Use a SVC classifier on the combined features
    ('svc', SVC(kernel='linear')),

# limit the list of categories to make running this example faster.
categories = ['alt.atheism', 'talk.religion.misc']
train = fetch_20newsgroups(random_state=1,
test = fetch_20newsgroups(random_state=1,
y = pipeline.predict(



    Matt Terry <>


    BSD 3 clause