Feature Union with Heterogeneous Data Sources

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.

# Author: Matt Terry <matt.terry@gmail.com>
#
# 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 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`.
 
    Parameters
    ----------
    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):]
                    break
            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(
        transformer_list=[
 
            # 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
        transformer_weights={
            '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,
                           subset='train',
                           categories=categories,
                           )
test = fetch_20newsgroups(random_state=1,
                          subset='test',
                          categories=categories,
                          )
 
pipeline.fit(train.data, train.target)
y = pipeline.predict(test.data)
print(classification_report(y, test.target))

Total running time of the script: (0 minutes 0.000 seconds)

Download Python source code: hetero_feature_union.py

Download IPython notebook: hetero_feature_union.ipynb

Links:

http://scikit-learn.org/stable/auto_examples/hetero_feature_union.html

doc_scikit_learn

2025-01-10 15:47:30

Comments