Fork me on GitHub

Feature Union with Heterogeneous Data SourcesΒΆ

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

  1. Your dataset consists of heterogeneous data types (e.g. raster images and text captions)
  2. 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.

Python source code:

# Author: Matt Terry <>
# License: BSD 3 clause
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

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 sklearn 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 exmaple faster.
categories = ['alt.atheism', 'talk.religion.misc']
train = fetch_20newsgroups(random_state=1,
test = fetch_20newsgroups(random_state=1,
y = pipeline.predict(