-
Notifications
You must be signed in to change notification settings - Fork 81
/
Copy pathfeature_extraction.html
1084 lines (1035 loc) · 100 KB
/
feature_extraction.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN"
"http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd">
<html xmlns="http://www.w3.org/1999/xhtml">
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8" />
<title>4.2. Feature extraction — scikit-learn 0.16.1 documentation</title>
<!-- htmltitle is before nature.css - we use this hack to load bootstrap first -->
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<link rel="stylesheet" href="../_static/css/bootstrap.min.css" media="screen" />
<link rel="stylesheet" href="../_static/css/bootstrap-responsive.css"/>
<link rel="stylesheet" href="../_static/nature.css" type="text/css" />
<link rel="stylesheet" href="../_static/pygments.css" type="text/css" />
<link rel="stylesheet" href="../_static/gallery.css" type="text/css" />
<script type="text/javascript">
var DOCUMENTATION_OPTIONS = {
URL_ROOT: '../',
VERSION: '0.16.1',
COLLAPSE_INDEX: false,
FILE_SUFFIX: '.html',
HAS_SOURCE: true
};
</script>
<script type="text/javascript" src="../_static/jquery.js"></script>
<script type="text/javascript" src="../_static/underscore.js"></script>
<script type="text/javascript" src="../_static/doctools.js"></script>
<script type="text/javascript" src="../_static/js/copybutton.js"></script>
<link rel="shortcut icon" href="../_static/favicon.ico"/>
<link rel="author" title="About these documents" href="../about.html" />
<link rel="top" title="scikit-learn 0.16.1 documentation" href="../index.html" />
<link rel="up" title="4. Dataset transformations" href="../data_transforms.html" />
<link rel="next" title="4.3. Preprocessing data" href="preprocessing.html" />
<link rel="prev" title="4.1. Pipeline and FeatureUnion: combining estimators" href="pipeline.html" />
<script type="text/javascript" src="../_static/sidebar.js"></script>
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<script src="../_static/js/bootstrap.min.js" type="text/javascript"></script>
<link rel="canonical" href="https://scikit-learn.org/stable/modules/feature_extraction.html" />
<script type="text/javascript">
$("div.buttonNext, div.buttonPrevious").hover(
function () {
$(this).css('background-color', '#FF9C34');
},
function () {
$(this).css('background-color', '#A7D6E2');
}
);
var bodywrapper = $('.bodywrapper');
var sidebarbutton = $('#sidebarbutton');
sidebarbutton.css({'height': '900px'});
</script>
</head>
<body>
<div class="header-wrapper">
<div class="header">
<p class="logo"><a href="../index.html">
<img src="../_static/scikit-learn-logo-small.png" alt="Logo"/>
</a>
</p><div class="navbar">
<ul>
<li><a href="../../stable/index.html">Home</a></li>
<li><a href="../../stable/install.html">Installation</a></li>
<li class="btn-li"><div class="btn-group">
<a href="../documentation.html">Documentation</a>
<a class="btn dropdown-toggle" data-toggle="dropdown">
<span class="caret"></span>
</a>
<ul class="dropdown-menu">
<li class="link-title">Scikit-learn 0.16 (Stable)</li>
<li><a href="../tutorial/index.html">Tutorials</a></li>
<li><a href="../user_guide.html">User guide</a></li>
<li><a href="classes.html">API</a></li>
<li><a href="../faq.html">FAQ</a></li>
<li class="divider"></li>
<li><a href="http://scikit-learn.org/dev/documentation.html">Development</a></li>
<li><a href="http://scikit-learn.org/0.15/">Scikit-learn 0.15</a></li>
</ul>
</div>
</li>
<li><a href="../auto_examples/index.html">Examples</a></li>
</ul>
<div class="search_form">
<div id="cse" style="width: 100%;"></div>
</div>
</div> <!-- end navbar --></div>
</div>
<!-- Github "fork me" ribbon -->
<a href="https://github.com/scikit-learn/scikit-learn">
<img class="fork-me"
style="position: absolute; top: 0; right: 0; border: 0;"
src="../_static/img/forkme.png"
alt="Fork me on GitHub" />
</a>
<div class="content-wrapper">
<div class="sphinxsidebar">
<div class="sphinxsidebarwrapper">
<div class="rel">
<!-- rellinks[1:] is an ugly hack to avoid link to module
index -->
<div class="rellink">
<a href="pipeline.html"
accesskey="P">Previous
<br/>
<span class="smallrellink">
4.1. Pipeline an...
</span>
<span class="hiddenrellink">
4.1. Pipeline and FeatureUnion: combining estimators
</span>
</a>
</div>
<div class="spacer">
</div>
<div class="rellink">
<a href="preprocessing.html"
accesskey="N">Next
<br/>
<span class="smallrellink">
4.3. Preprocessi...
</span>
<span class="hiddenrellink">
4.3. Preprocessing data
</span>
</a>
</div>
<!-- Ad a link to the 'up' page -->
<div class="spacer">
</div>
<div class="rellink">
<a href="../data_transforms.html">
Up
<br/>
<span class="smallrellink">
4. Dataset trans...
</span>
<span class="hiddenrellink">
4. Dataset transformations
</span>
</a>
</div>
</div>
<p class="doc-version">This documentation is for scikit-learn <strong>version 0.16.1</strong> — <a href="http://scikit-learn.org/stable/support.html#documentation-resources">Other versions</a></p>
<p class="citing">If you use the software, please consider <a href="../about.html#citing-scikit-learn">citing scikit-learn</a>.</p>
<ul>
<li><a class="reference internal" href="#">4.2. Feature extraction</a><ul>
<li><a class="reference internal" href="#loading-features-from-dicts">4.2.1. Loading features from dicts</a></li>
<li><a class="reference internal" href="#feature-hashing">4.2.2. Feature hashing</a><ul>
<li><a class="reference internal" href="#implementation-details">4.2.2.1. Implementation details</a></li>
</ul>
</li>
<li><a class="reference internal" href="#text-feature-extraction">4.2.3. Text feature extraction</a><ul>
<li><a class="reference internal" href="#the-bag-of-words-representation">4.2.3.1. The Bag of Words representation</a></li>
<li><a class="reference internal" href="#sparsity">4.2.3.2. Sparsity</a></li>
<li><a class="reference internal" href="#common-vectorizer-usage">4.2.3.3. Common Vectorizer usage</a></li>
<li><a class="reference internal" href="#tfidf-term-weighting">4.2.3.4. Tf–idf term weighting</a></li>
<li><a class="reference internal" href="#decoding-text-files">4.2.3.5. Decoding text files</a></li>
<li><a class="reference internal" href="#applications-and-examples">4.2.3.6. Applications and examples</a></li>
<li><a class="reference internal" href="#limitations-of-the-bag-of-words-representation">4.2.3.7. Limitations of the Bag of Words representation</a></li>
<li><a class="reference internal" href="#vectorizing-a-large-text-corpus-with-the-hashing-trick">4.2.3.8. Vectorizing a large text corpus with the hashing trick</a></li>
<li><a class="reference internal" href="#performing-out-of-core-scaling-with-hashingvectorizer">4.2.3.9. Performing out-of-core scaling with HashingVectorizer</a></li>
<li><a class="reference internal" href="#customizing-the-vectorizer-classes">4.2.3.10. Customizing the vectorizer classes</a></li>
</ul>
</li>
<li><a class="reference internal" href="#image-feature-extraction">4.2.4. Image feature extraction</a><ul>
<li><a class="reference internal" href="#patch-extraction">4.2.4.1. Patch extraction</a></li>
<li><a class="reference internal" href="#connectivity-graph-of-an-image">4.2.4.2. Connectivity graph of an image</a></li>
</ul>
</li>
</ul>
</li>
</ul>
</div>
</div>
<div class="content">
<div class="documentwrapper">
<div class="bodywrapper">
<div class="body">
<div class="section" id="feature-extraction">
<span id="id1"></span><h1>4.2. Feature extraction<a class="headerlink" href="#feature-extraction" title="Permalink to this headline">¶</a></h1>
<p>The <a class="reference internal" href="classes.html#module-sklearn.feature_extraction" title="sklearn.feature_extraction"><tt class="xref py py-mod docutils literal"><span class="pre">sklearn.feature_extraction</span></tt></a> module can be used to extract
features in a format supported by machine learning algorithms from datasets
consisting of formats such as text and image.</p>
<div class="admonition note">
<p class="first admonition-title">Note</p>
<p class="last">Feature extraction is very different from <a class="reference internal" href="feature_selection.html#feature-selection"><em>Feature selection</em></a>:
the former consists in transforming arbitrary data, such as text or
images, into numerical features usable for machine learning. The latter
is a machine learning technique applied on these features.</p>
</div>
<div class="section" id="loading-features-from-dicts">
<span id="dict-feature-extraction"></span><h2>4.2.1. Loading features from dicts<a class="headerlink" href="#loading-features-from-dicts" title="Permalink to this headline">¶</a></h2>
<p>The class <a class="reference internal" href="generated/sklearn.feature_extraction.DictVectorizer.html#sklearn.feature_extraction.DictVectorizer" title="sklearn.feature_extraction.DictVectorizer"><tt class="xref py py-class docutils literal"><span class="pre">DictVectorizer</span></tt></a> can be used to convert feature
arrays represented as lists of standard Python <tt class="docutils literal"><span class="pre">dict</span></tt> objects to the
NumPy/SciPy representation used by scikit-learn estimators.</p>
<p>While not particularly fast to process, Python’s <tt class="docutils literal"><span class="pre">dict</span></tt> has the
advantages of being convenient to use, being sparse (absent features
need not be stored) and storing feature names in addition to values.</p>
<p><a class="reference internal" href="generated/sklearn.feature_extraction.DictVectorizer.html#sklearn.feature_extraction.DictVectorizer" title="sklearn.feature_extraction.DictVectorizer"><tt class="xref py py-class docutils literal"><span class="pre">DictVectorizer</span></tt></a> implements what is called one-of-K or “one-hot”
coding for categorical (aka nominal, discrete) features. Categorical
features are “attribute-value” pairs where the value is restricted
to a list of discrete of possibilities without ordering (e.g. topic
identifiers, types of objects, tags, names...).</p>
<p>In the following, “city” is a categorical attribute while “temperature”
is a traditional numerical feature:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="n">measurements</span> <span class="o">=</span> <span class="p">[</span>
<span class="gp">... </span> <span class="p">{</span><span class="s">'city'</span><span class="p">:</span> <span class="s">'Dubai'</span><span class="p">,</span> <span class="s">'temperature'</span><span class="p">:</span> <span class="mf">33.</span><span class="p">},</span>
<span class="gp">... </span> <span class="p">{</span><span class="s">'city'</span><span class="p">:</span> <span class="s">'London'</span><span class="p">,</span> <span class="s">'temperature'</span><span class="p">:</span> <span class="mf">12.</span><span class="p">},</span>
<span class="gp">... </span> <span class="p">{</span><span class="s">'city'</span><span class="p">:</span> <span class="s">'San Fransisco'</span><span class="p">,</span> <span class="s">'temperature'</span><span class="p">:</span> <span class="mf">18.</span><span class="p">},</span>
<span class="gp">... </span><span class="p">]</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">sklearn.feature_extraction</span> <span class="kn">import</span> <span class="n">DictVectorizer</span>
<span class="gp">>>> </span><span class="n">vec</span> <span class="o">=</span> <span class="n">DictVectorizer</span><span class="p">()</span>
<span class="gp">>>> </span><span class="n">vec</span><span class="o">.</span><span class="n">fit_transform</span><span class="p">(</span><span class="n">measurements</span><span class="p">)</span><span class="o">.</span><span class="n">toarray</span><span class="p">()</span>
<span class="go">array([[ 1., 0., 0., 33.],</span>
<span class="go"> [ 0., 1., 0., 12.],</span>
<span class="go"> [ 0., 0., 1., 18.]])</span>
<span class="gp">>>> </span><span class="n">vec</span><span class="o">.</span><span class="n">get_feature_names</span><span class="p">()</span>
<span class="go">['city=Dubai', 'city=London', 'city=San Fransisco', 'temperature']</span>
</pre></div>
</div>
<p><a class="reference internal" href="generated/sklearn.feature_extraction.DictVectorizer.html#sklearn.feature_extraction.DictVectorizer" title="sklearn.feature_extraction.DictVectorizer"><tt class="xref py py-class docutils literal"><span class="pre">DictVectorizer</span></tt></a> is also a useful representation transformation
for training sequence classifiers in Natural Language Processing models
that typically work by extracting feature windows around a particular
word of interest.</p>
<p>For example, suppose that we have a first algorithm that extracts Part of
Speech (PoS) tags that we want to use as complementary tags for training
a sequence classifier (e.g. a chunker). The following dict could be
such a window of features extracted around the word ‘sat’ in the sentence
‘The cat sat on the mat.’:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="n">pos_window</span> <span class="o">=</span> <span class="p">[</span>
<span class="gp">... </span> <span class="p">{</span>
<span class="gp">... </span> <span class="s">'word-2'</span><span class="p">:</span> <span class="s">'the'</span><span class="p">,</span>
<span class="gp">... </span> <span class="s">'pos-2'</span><span class="p">:</span> <span class="s">'DT'</span><span class="p">,</span>
<span class="gp">... </span> <span class="s">'word-1'</span><span class="p">:</span> <span class="s">'cat'</span><span class="p">,</span>
<span class="gp">... </span> <span class="s">'pos-1'</span><span class="p">:</span> <span class="s">'NN'</span><span class="p">,</span>
<span class="gp">... </span> <span class="s">'word+1'</span><span class="p">:</span> <span class="s">'on'</span><span class="p">,</span>
<span class="gp">... </span> <span class="s">'pos+1'</span><span class="p">:</span> <span class="s">'PP'</span><span class="p">,</span>
<span class="gp">... </span> <span class="p">},</span>
<span class="gp">... </span> <span class="c"># in a real application one would extract many such dictionaries</span>
<span class="gp">... </span><span class="p">]</span>
</pre></div>
</div>
<p>This description can be vectorized into a sparse two-dimensional matrix
suitable for feeding into a classifier (maybe after being piped into a
<a class="reference internal" href="generated/sklearn.feature_extraction.text.TfidfTransformer.html#sklearn.feature_extraction.text.TfidfTransformer" title="sklearn.feature_extraction.text.TfidfTransformer"><tt class="xref py py-class docutils literal"><span class="pre">text.TfidfTransformer</span></tt></a> for normalization):</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="n">vec</span> <span class="o">=</span> <span class="n">DictVectorizer</span><span class="p">()</span>
<span class="gp">>>> </span><span class="n">pos_vectorized</span> <span class="o">=</span> <span class="n">vec</span><span class="o">.</span><span class="n">fit_transform</span><span class="p">(</span><span class="n">pos_window</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">pos_vectorized</span>
<span class="go"><1x6 sparse matrix of type '<... 'numpy.float64'>'</span>
<span class="go"> with 6 stored elements in Compressed Sparse ... format></span>
<span class="gp">>>> </span><span class="n">pos_vectorized</span><span class="o">.</span><span class="n">toarray</span><span class="p">()</span>
<span class="go">array([[ 1., 1., 1., 1., 1., 1.]])</span>
<span class="gp">>>> </span><span class="n">vec</span><span class="o">.</span><span class="n">get_feature_names</span><span class="p">()</span>
<span class="go">['pos+1=PP', 'pos-1=NN', 'pos-2=DT', 'word+1=on', 'word-1=cat', 'word-2=the']</span>
</pre></div>
</div>
<p>As you can imagine, if one extracts such a context around each individual
word of a corpus of documents the resulting matrix will be very wide
(many one-hot-features) with most of them being valued to zero most
of the time. So as to make the resulting data structure able to fit in
memory the <tt class="docutils literal"><span class="pre">DictVectorizer</span></tt> class uses a <tt class="docutils literal"><span class="pre">scipy.sparse</span></tt> matrix by
default instead of a <tt class="docutils literal"><span class="pre">numpy.ndarray</span></tt>.</p>
</div>
<div class="section" id="feature-hashing">
<span id="id2"></span><h2>4.2.2. Feature hashing<a class="headerlink" href="#feature-hashing" title="Permalink to this headline">¶</a></h2>
<p>The class <a class="reference internal" href="generated/sklearn.feature_extraction.FeatureHasher.html#sklearn.feature_extraction.FeatureHasher" title="sklearn.feature_extraction.FeatureHasher"><tt class="xref py py-class docutils literal"><span class="pre">FeatureHasher</span></tt></a> is a high-speed, low-memory vectorizer that
uses a technique known as
<a class="reference external" href="https://en.wikipedia.org/wiki/Feature_hashing">feature hashing</a>,
or the “hashing trick”.
Instead of building a hash table of the features encountered in training,
as the vectorizers do, instances of <a class="reference internal" href="generated/sklearn.feature_extraction.FeatureHasher.html#sklearn.feature_extraction.FeatureHasher" title="sklearn.feature_extraction.FeatureHasher"><tt class="xref py py-class docutils literal"><span class="pre">FeatureHasher</span></tt></a>
apply a hash function to the features
to determine their column index in sample matrices directly.
The result is increased speed and reduced memory usage,
at the expense of inspectability;
the hasher does not remember what the input features looked like
and has no <tt class="docutils literal"><span class="pre">inverse_transform</span></tt> method.</p>
<p>Since the hash function might cause collisions between (unrelated) features,
a signed hash function is used and the sign of the hash value
determines the sign of the value stored in the output matrix for a feature.
This way, collisions are likely to cancel out rather than accumulate error,
and the expected mean of any output feature’s value is zero.</p>
<p>If <tt class="docutils literal"><span class="pre">non_negative=True</span></tt> is passed to the constructor, the absolute
value is taken. This undoes some of the collision handling, but allows
the output to be passed to estimators like
<a class="reference internal" href="generated/sklearn.naive_bayes.MultinomialNB.html#sklearn.naive_bayes.MultinomialNB" title="sklearn.naive_bayes.MultinomialNB"><tt class="xref py py-class docutils literal"><span class="pre">sklearn.naive_bayes.MultinomialNB</span></tt></a> or
<a class="reference internal" href="generated/sklearn.feature_selection.chi2.html#sklearn.feature_selection.chi2" title="sklearn.feature_selection.chi2"><tt class="xref py py-class docutils literal"><span class="pre">sklearn.feature_selection.chi2</span></tt></a>
feature selectors that expect non-negative inputs.</p>
<p><a class="reference internal" href="generated/sklearn.feature_extraction.FeatureHasher.html#sklearn.feature_extraction.FeatureHasher" title="sklearn.feature_extraction.FeatureHasher"><tt class="xref py py-class docutils literal"><span class="pre">FeatureHasher</span></tt></a> accepts either mappings
(like Python’s <tt class="docutils literal"><span class="pre">dict</span></tt> and its variants in the <tt class="docutils literal"><span class="pre">collections</span></tt> module),
<tt class="docutils literal"><span class="pre">(feature,</span> <span class="pre">value)</span></tt> pairs, or strings,
depending on the constructor parameter <tt class="docutils literal"><span class="pre">input_type</span></tt>.
Mapping are treated as lists of <tt class="docutils literal"><span class="pre">(feature,</span> <span class="pre">value)</span></tt> pairs,
while single strings have an implicit value of 1,
so <tt class="docutils literal"><span class="pre">['feat1',</span> <span class="pre">'feat2',</span> <span class="pre">'feat3']</span></tt> is interpreted as
<tt class="docutils literal"><span class="pre">[('feat1',</span> <span class="pre">1),</span> <span class="pre">('feat2',</span> <span class="pre">1),</span> <span class="pre">('feat3',</span> <span class="pre">1)]</span></tt>.
If a single feature occurs multiple times in a sample,
the associated values will be summed
(so <tt class="docutils literal"><span class="pre">('feat',</span> <span class="pre">2)</span></tt> and <tt class="docutils literal"><span class="pre">('feat',</span> <span class="pre">3.5)</span></tt> become <tt class="docutils literal"><span class="pre">('feat',</span> <span class="pre">5.5)</span></tt>).
The output from <a class="reference internal" href="generated/sklearn.feature_extraction.FeatureHasher.html#sklearn.feature_extraction.FeatureHasher" title="sklearn.feature_extraction.FeatureHasher"><tt class="xref py py-class docutils literal"><span class="pre">FeatureHasher</span></tt></a> is always a <tt class="docutils literal"><span class="pre">scipy.sparse</span></tt> matrix
in the CSR format.</p>
<p>Feature hashing can be employed in document classification,
but unlike <a class="reference internal" href="generated/sklearn.feature_extraction.text.CountVectorizer.html#sklearn.feature_extraction.text.CountVectorizer" title="sklearn.feature_extraction.text.CountVectorizer"><tt class="xref py py-class docutils literal"><span class="pre">text.CountVectorizer</span></tt></a>,
<a class="reference internal" href="generated/sklearn.feature_extraction.FeatureHasher.html#sklearn.feature_extraction.FeatureHasher" title="sklearn.feature_extraction.FeatureHasher"><tt class="xref py py-class docutils literal"><span class="pre">FeatureHasher</span></tt></a> does not do word
splitting or any other preprocessing except Unicode-to-UTF-8 encoding;
see <a class="reference internal" href="#hashing-vectorizer"><em>Vectorizing a large text corpus with the hashing trick</em></a>, below, for a combined tokenizer/hasher.</p>
<p>As an example, consider a word-level natural language processing task
that needs features extracted from <tt class="docutils literal"><span class="pre">(token,</span> <span class="pre">part_of_speech)</span></tt> pairs.
One could use a Python generator function to extract features:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="k">def</span> <span class="nf">token_features</span><span class="p">(</span><span class="n">token</span><span class="p">,</span> <span class="n">part_of_speech</span><span class="p">):</span>
<span class="k">if</span> <span class="n">token</span><span class="o">.</span><span class="n">isdigit</span><span class="p">():</span>
<span class="k">yield</span> <span class="s">"numeric"</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">yield</span> <span class="s">"token={}"</span><span class="o">.</span><span class="n">format</span><span class="p">(</span><span class="n">token</span><span class="o">.</span><span class="n">lower</span><span class="p">())</span>
<span class="k">yield</span> <span class="s">"token,pos={},{}"</span><span class="o">.</span><span class="n">format</span><span class="p">(</span><span class="n">token</span><span class="p">,</span> <span class="n">part_of_speech</span><span class="p">)</span>
<span class="k">if</span> <span class="n">token</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span><span class="o">.</span><span class="n">isupper</span><span class="p">():</span>
<span class="k">yield</span> <span class="s">"uppercase_initial"</span>
<span class="k">if</span> <span class="n">token</span><span class="o">.</span><span class="n">isupper</span><span class="p">():</span>
<span class="k">yield</span> <span class="s">"all_uppercase"</span>
<span class="k">yield</span> <span class="s">"pos={}"</span><span class="o">.</span><span class="n">format</span><span class="p">(</span><span class="n">part_of_speech</span><span class="p">)</span>
</pre></div>
</div>
<p>Then, the <tt class="docutils literal"><span class="pre">raw_X</span></tt> to be fed to <tt class="docutils literal"><span class="pre">FeatureHasher.transform</span></tt>
can be constructed using:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="n">raw_X</span> <span class="o">=</span> <span class="p">(</span><span class="n">token_features</span><span class="p">(</span><span class="n">tok</span><span class="p">,</span> <span class="n">pos_tagger</span><span class="p">(</span><span class="n">tok</span><span class="p">))</span> <span class="k">for</span> <span class="n">tok</span> <span class="ow">in</span> <span class="n">corpus</span><span class="p">)</span>
</pre></div>
</div>
<p>and fed to a hasher with:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="n">hasher</span> <span class="o">=</span> <span class="n">FeatureHasher</span><span class="p">(</span><span class="n">input_type</span><span class="o">=</span><span class="s">'string'</span><span class="p">)</span>
<span class="n">X</span> <span class="o">=</span> <span class="n">hasher</span><span class="o">.</span><span class="n">transform</span><span class="p">(</span><span class="n">raw_X</span><span class="p">)</span>
</pre></div>
</div>
<p>to get a <tt class="docutils literal"><span class="pre">scipy.sparse</span></tt> matrix <tt class="docutils literal"><span class="pre">X</span></tt>.</p>
<p>Note the use of a generator comprehension,
which introduces laziness into the feature extraction:
tokens are only processed on demand from the hasher.</p>
<div class="section" id="implementation-details">
<h3>4.2.2.1. Implementation details<a class="headerlink" href="#implementation-details" title="Permalink to this headline">¶</a></h3>
<p><a class="reference internal" href="generated/sklearn.feature_extraction.FeatureHasher.html#sklearn.feature_extraction.FeatureHasher" title="sklearn.feature_extraction.FeatureHasher"><tt class="xref py py-class docutils literal"><span class="pre">FeatureHasher</span></tt></a> uses the signed 32-bit variant of MurmurHash3.
As a result (and because of limitations in <tt class="docutils literal"><span class="pre">scipy.sparse</span></tt>),
the maximum number of features supported is currently <img class="math" src="../_images/math/93c8c4826c6f7713528b0a263ab8b95d736366c6.png" alt="2^{31} - 1"/>.</p>
<p>The original formulation of the hashing trick by Weinberger et al.
used two separate hash functions <img class="math" src="../_images/math/cbb80ad77aa7a5e227d5a46bc44d235284106cfc.png" alt="h"/> and <img class="math" src="../_images/math/26bfd3bc4c5db78139d8897f9cfa5bba511a2f22.png" alt="\xi"/>
to determine the column index and sign of a feature, respectively.
The present implementation works under the assumption
that the sign bit of MurmurHash3 is independent of its other bits.</p>
<p>Since a simple modulo is used to transform the hash function to a column index,
it is advisable to use a power of two as the <tt class="docutils literal"><span class="pre">n_features</span></tt> parameter;
otherwise the features will not be mapped evenly to the columns.</p>
<div class="topic">
<p class="topic-title first">References:</p>
<ul class="simple">
<li>Kilian Weinberger, Anirban Dasgupta, John Langford, Alex Smola and
Josh Attenberg (2009). <a class="reference external" href="http://alex.smola.org/papers/2009/Weinbergeretal09.pdf">Feature hashing for large scale multitask learning</a>. Proc. ICML.</li>
<li><a class="reference external" href="http://code.google.com/p/smhasher/wiki/MurmurHash3">MurmurHash3</a>.</li>
</ul>
</div>
</div>
</div>
<div class="section" id="text-feature-extraction">
<span id="id4"></span><h2>4.2.3. Text feature extraction<a class="headerlink" href="#text-feature-extraction" title="Permalink to this headline">¶</a></h2>
<div class="section" id="the-bag-of-words-representation">
<h3>4.2.3.1. The Bag of Words representation<a class="headerlink" href="#the-bag-of-words-representation" title="Permalink to this headline">¶</a></h3>
<p>Text Analysis is a major application field for machine learning
algorithms. However the raw data, a sequence of symbols cannot be fed
directly to the algorithms themselves as most of them expect numerical
feature vectors with a fixed size rather than the raw text documents
with variable length.</p>
<p>In order to address this, scikit-learn provides utilities for the most
common ways to extract numerical features from text content, namely:</p>
<ul class="simple">
<li><strong>tokenizing</strong> strings and giving an integer id for each possible token,
for instance by using white-spaces and punctuation as token separators.</li>
<li><strong>counting</strong> the occurrences of tokens in each document.</li>
<li><strong>normalizing</strong> and weighting with diminishing importance tokens that
occur in the majority of samples / documents.</li>
</ul>
<p>In this scheme, features and samples are defined as follows:</p>
<ul class="simple">
<li>each <strong>individual token occurrence frequency</strong> (normalized or not)
is treated as a <strong>feature</strong>.</li>
<li>the vector of all the token frequencies for a given <strong>document</strong> is
considered a multivariate <strong>sample</strong>.</li>
</ul>
<p>A corpus of documents can thus be represented by a matrix with one row
per document and one column per token (e.g. word) occurring in the corpus.</p>
<p>We call <strong>vectorization</strong> the general process of turning a collection
of text documents into numerical feature vectors. This specific strategy
(tokenization, counting and normalization) is called the <strong>Bag of Words</strong>
or “Bag of n-grams” representation. Documents are described by word
occurrences while completely ignoring the relative position information
of the words in the document.</p>
</div>
<div class="section" id="sparsity">
<h3>4.2.3.2. Sparsity<a class="headerlink" href="#sparsity" title="Permalink to this headline">¶</a></h3>
<p>As most documents will typically use a very small subset of the words used in
the corpus, the resulting matrix will have many feature values that are
zeros (typically more than 99% of them).</p>
<p>For instance a collection of 10,000 short text documents (such as emails)
will use a vocabulary with a size in the order of 100,000 unique words in
total while each document will use 100 to 1000 unique words individually.</p>
<p>In order to be able to store such a matrix in memory but also to speed
up algebraic operations matrix / vector, implementations will typically
use a sparse representation such as the implementations available in the
<tt class="docutils literal"><span class="pre">scipy.sparse</span></tt> package.</p>
</div>
<div class="section" id="common-vectorizer-usage">
<h3>4.2.3.3. Common Vectorizer usage<a class="headerlink" href="#common-vectorizer-usage" title="Permalink to this headline">¶</a></h3>
<p><a class="reference internal" href="generated/sklearn.feature_extraction.text.CountVectorizer.html#sklearn.feature_extraction.text.CountVectorizer" title="sklearn.feature_extraction.text.CountVectorizer"><tt class="xref py py-class docutils literal"><span class="pre">CountVectorizer</span></tt></a> implements both tokenization and occurrence
counting in a single class:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">sklearn.feature_extraction.text</span> <span class="kn">import</span> <span class="n">CountVectorizer</span>
</pre></div>
</div>
<p>This model has many parameters, however the default values are quite
reasonable (please see the <a class="reference internal" href="classes.html#text-feature-extraction-ref"><em>reference documentation</em></a> for the details):</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="n">vectorizer</span> <span class="o">=</span> <span class="n">CountVectorizer</span><span class="p">(</span><span class="n">min_df</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">vectorizer</span>
<span class="go">CountVectorizer(analyzer=...'word', binary=False, decode_error=...'strict',</span>
<span class="go"> dtype=<... 'numpy.int64'>, encoding=...'utf-8', input=...'content',</span>
<span class="go"> lowercase=True, max_df=1.0, max_features=None, min_df=1,</span>
<span class="go"> ngram_range=(1, 1), preprocessor=None, stop_words=None,</span>
<span class="go"> strip_accents=None, token_pattern=...'(?u)\\b\\w\\w+\\b',</span>
<span class="go"> tokenizer=None, vocabulary=None)</span>
</pre></div>
</div>
<p>Let’s use it to tokenize and count the word occurrences of a minimalistic
corpus of text documents:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="n">corpus</span> <span class="o">=</span> <span class="p">[</span>
<span class="gp">... </span> <span class="s">'This is the first document.'</span><span class="p">,</span>
<span class="gp">... </span> <span class="s">'This is the second second document.'</span><span class="p">,</span>
<span class="gp">... </span> <span class="s">'And the third one.'</span><span class="p">,</span>
<span class="gp">... </span> <span class="s">'Is this the first document?'</span><span class="p">,</span>
<span class="gp">... </span><span class="p">]</span>
<span class="gp">>>> </span><span class="n">X</span> <span class="o">=</span> <span class="n">vectorizer</span><span class="o">.</span><span class="n">fit_transform</span><span class="p">(</span><span class="n">corpus</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">X</span>
<span class="go"><4x9 sparse matrix of type '<... 'numpy.int64'>'</span>
<span class="go"> with 19 stored elements in Compressed Sparse ... format></span>
</pre></div>
</div>
<p>The default configuration tokenizes the string by extracting words of
at least 2 letters. The specific function that does this step can be
requested explicitly:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="n">analyze</span> <span class="o">=</span> <span class="n">vectorizer</span><span class="o">.</span><span class="n">build_analyzer</span><span class="p">()</span>
<span class="gp">>>> </span><span class="n">analyze</span><span class="p">(</span><span class="s">"This is a text document to analyze."</span><span class="p">)</span> <span class="o">==</span> <span class="p">(</span>
<span class="gp">... </span> <span class="p">[</span><span class="s">'this'</span><span class="p">,</span> <span class="s">'is'</span><span class="p">,</span> <span class="s">'text'</span><span class="p">,</span> <span class="s">'document'</span><span class="p">,</span> <span class="s">'to'</span><span class="p">,</span> <span class="s">'analyze'</span><span class="p">])</span>
<span class="go">True</span>
</pre></div>
</div>
<p>Each term found by the analyzer during the fit is assigned a unique
integer index corresponding to a column in the resulting matrix. This
interpretation of the columns can be retrieved as follows:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="n">vectorizer</span><span class="o">.</span><span class="n">get_feature_names</span><span class="p">()</span> <span class="o">==</span> <span class="p">(</span>
<span class="gp">... </span> <span class="p">[</span><span class="s">'and'</span><span class="p">,</span> <span class="s">'document'</span><span class="p">,</span> <span class="s">'first'</span><span class="p">,</span> <span class="s">'is'</span><span class="p">,</span> <span class="s">'one'</span><span class="p">,</span>
<span class="gp">... </span> <span class="s">'second'</span><span class="p">,</span> <span class="s">'the'</span><span class="p">,</span> <span class="s">'third'</span><span class="p">,</span> <span class="s">'this'</span><span class="p">])</span>
<span class="go">True</span>
<span class="gp">>>> </span><span class="n">X</span><span class="o">.</span><span class="n">toarray</span><span class="p">()</span>
<span class="go">array([[0, 1, 1, 1, 0, 0, 1, 0, 1],</span>
<span class="go"> [0, 1, 0, 1, 0, 2, 1, 0, 1],</span>
<span class="go"> [1, 0, 0, 0, 1, 0, 1, 1, 0],</span>
<span class="go"> [0, 1, 1, 1, 0, 0, 1, 0, 1]]...)</span>
</pre></div>
</div>
<p>The converse mapping from feature name to column index is stored in the
<tt class="docutils literal"><span class="pre">vocabulary_</span></tt> attribute of the vectorizer:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="n">vectorizer</span><span class="o">.</span><span class="n">vocabulary_</span><span class="o">.</span><span class="n">get</span><span class="p">(</span><span class="s">'document'</span><span class="p">)</span>
<span class="go">1</span>
</pre></div>
</div>
<p>Hence words that were not seen in the training corpus will be completely
ignored in future calls to the transform method:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="n">vectorizer</span><span class="o">.</span><span class="n">transform</span><span class="p">([</span><span class="s">'Something completely new.'</span><span class="p">])</span><span class="o">.</span><span class="n">toarray</span><span class="p">()</span>
<span class="gp">... </span>
<span class="go">array([[0, 0, 0, 0, 0, 0, 0, 0, 0]]...)</span>
</pre></div>
</div>
<p>Note that in the previous corpus, the first and the last documents have
exactly the same words hence are encoded in equal vectors. In particular
we lose the information that the last document is an interrogative form. To
preserve some of the local ordering information we can extract 2-grams
of words in addition to the 1-grams (individual words):</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="n">bigram_vectorizer</span> <span class="o">=</span> <span class="n">CountVectorizer</span><span class="p">(</span><span class="n">ngram_range</span><span class="o">=</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="mi">2</span><span class="p">),</span>
<span class="gp">... </span> <span class="n">token_pattern</span><span class="o">=</span><span class="s">r'\b\w+\b'</span><span class="p">,</span> <span class="n">min_df</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">analyze</span> <span class="o">=</span> <span class="n">bigram_vectorizer</span><span class="o">.</span><span class="n">build_analyzer</span><span class="p">()</span>
<span class="gp">>>> </span><span class="n">analyze</span><span class="p">(</span><span class="s">'Bi-grams are cool!'</span><span class="p">)</span> <span class="o">==</span> <span class="p">(</span>
<span class="gp">... </span> <span class="p">[</span><span class="s">'bi'</span><span class="p">,</span> <span class="s">'grams'</span><span class="p">,</span> <span class="s">'are'</span><span class="p">,</span> <span class="s">'cool'</span><span class="p">,</span> <span class="s">'bi grams'</span><span class="p">,</span> <span class="s">'grams are'</span><span class="p">,</span> <span class="s">'are cool'</span><span class="p">])</span>
<span class="go">True</span>
</pre></div>
</div>
<p>The vocabulary extracted by this vectorizer is hence much bigger and
can now resolve ambiguities encoded in local positioning patterns:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="n">X_2</span> <span class="o">=</span> <span class="n">bigram_vectorizer</span><span class="o">.</span><span class="n">fit_transform</span><span class="p">(</span><span class="n">corpus</span><span class="p">)</span><span class="o">.</span><span class="n">toarray</span><span class="p">()</span>
<span class="gp">>>> </span><span class="n">X_2</span>
<span class="gp">... </span>
<span class="go">array([[0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 1, 1, 0],</span>
<span class="go"> [0, 0, 1, 0, 0, 1, 1, 0, 0, 2, 1, 1, 1, 0, 1, 0, 0, 0, 1, 1, 0],</span>
<span class="go"> [1, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 1, 1, 1, 0, 0, 0],</span>
<span class="go"> [0, 0, 1, 1, 1, 1, 0, 1, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 1, 0, 1]]...)</span>
</pre></div>
</div>
<p>In particular the interrogative form “Is this” is only present in the
last document:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="n">feature_index</span> <span class="o">=</span> <span class="n">bigram_vectorizer</span><span class="o">.</span><span class="n">vocabulary_</span><span class="o">.</span><span class="n">get</span><span class="p">(</span><span class="s">'is this'</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">X_2</span><span class="p">[:,</span> <span class="n">feature_index</span><span class="p">]</span>
<span class="go">array([0, 0, 0, 1]...)</span>
</pre></div>
</div>
</div>
<div class="section" id="tfidf-term-weighting">
<span id="tfidf"></span><h3>4.2.3.4. Tf–idf term weighting<a class="headerlink" href="#tfidf-term-weighting" title="Permalink to this headline">¶</a></h3>
<p>In a large text corpus, some words will be very present (e.g. “the”, “a”,
“is” in English) hence carrying very little meaningful information about
the actual contents of the document. If we were to feed the direct count
data directly to a classifier those very frequent terms would shadow
the frequencies of rarer yet more interesting terms.</p>
<p>In order to re-weight the count features into floating point values
suitable for usage by a classifier it is very common to use the tf–idf
transform.</p>
<p>Tf means <strong>term-frequency</strong> while tf–idf means term-frequency times
<strong>inverse document-frequency</strong>. This is a originally a term weighting
scheme developed for information retrieval (as a ranking function
for search engines results), that has also found good use in document
classification and clustering.</p>
<p>This normalization is implemented by the <tt class="xref py py-class docutils literal"><span class="pre">text.TfidfTransformer</span></tt>
class:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">sklearn.feature_extraction.text</span> <span class="kn">import</span> <span class="n">TfidfTransformer</span>
<span class="gp">>>> </span><span class="n">transformer</span> <span class="o">=</span> <span class="n">TfidfTransformer</span><span class="p">()</span>
<span class="gp">>>> </span><span class="n">transformer</span>
<span class="go">TfidfTransformer(norm=...'l2', smooth_idf=True, sublinear_tf=False,</span>
<span class="go"> use_idf=True)</span>
</pre></div>
</div>
<p>Again please see the <a class="reference internal" href="classes.html#text-feature-extraction-ref"><em>reference documentation</em></a> for the details on all the parameters.</p>
<p>Let’s take an example with the following counts. The first term is present
100% of the time hence not very interesting. The two other features only
in less than 50% of the time hence probably more representative of the
content of the documents:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="n">counts</span> <span class="o">=</span> <span class="p">[[</span><span class="mi">3</span><span class="p">,</span> <span class="mi">0</span><span class="p">,</span> <span class="mi">1</span><span class="p">],</span>
<span class="gp">... </span> <span class="p">[</span><span class="mi">2</span><span class="p">,</span> <span class="mi">0</span><span class="p">,</span> <span class="mi">0</span><span class="p">],</span>
<span class="gp">... </span> <span class="p">[</span><span class="mi">3</span><span class="p">,</span> <span class="mi">0</span><span class="p">,</span> <span class="mi">0</span><span class="p">],</span>
<span class="gp">... </span> <span class="p">[</span><span class="mi">4</span><span class="p">,</span> <span class="mi">0</span><span class="p">,</span> <span class="mi">0</span><span class="p">],</span>
<span class="gp">... </span> <span class="p">[</span><span class="mi">3</span><span class="p">,</span> <span class="mi">2</span><span class="p">,</span> <span class="mi">0</span><span class="p">],</span>
<span class="gp">... </span> <span class="p">[</span><span class="mi">3</span><span class="p">,</span> <span class="mi">0</span><span class="p">,</span> <span class="mi">2</span><span class="p">]]</span>
<span class="gp">...</span>
<span class="gp">>>> </span><span class="n">tfidf</span> <span class="o">=</span> <span class="n">transformer</span><span class="o">.</span><span class="n">fit_transform</span><span class="p">(</span><span class="n">counts</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">tfidf</span>
<span class="go"><6x3 sparse matrix of type '<... 'numpy.float64'>'</span>
<span class="go"> with 9 stored elements in Compressed Sparse ... format></span>
<span class="gp">>>> </span><span class="n">tfidf</span><span class="o">.</span><span class="n">toarray</span><span class="p">()</span>
<span class="go">array([[ 0.85..., 0. ..., 0.52...],</span>
<span class="go"> [ 1. ..., 0. ..., 0. ...],</span>
<span class="go"> [ 1. ..., 0. ..., 0. ...],</span>
<span class="go"> [ 1. ..., 0. ..., 0. ...],</span>
<span class="go"> [ 0.55..., 0.83..., 0. ...],</span>
<span class="go"> [ 0.63..., 0. ..., 0.77...]])</span>
</pre></div>
</div>
<p>Each row is normalized to have unit euclidean norm. The weights of each
feature computed by the <tt class="docutils literal"><span class="pre">fit</span></tt> method call are stored in a model
attribute:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="n">transformer</span><span class="o">.</span><span class="n">idf_</span>
<span class="go">array([ 1. ..., 2.25..., 1.84...])</span>
</pre></div>
</div>
<p>As tf–idf is very often used for text features, there is also another
class called <a class="reference internal" href="generated/sklearn.feature_extraction.text.TfidfVectorizer.html#sklearn.feature_extraction.text.TfidfVectorizer" title="sklearn.feature_extraction.text.TfidfVectorizer"><tt class="xref py py-class docutils literal"><span class="pre">TfidfVectorizer</span></tt></a> that combines all the options of
<a class="reference internal" href="generated/sklearn.feature_extraction.text.CountVectorizer.html#sklearn.feature_extraction.text.CountVectorizer" title="sklearn.feature_extraction.text.CountVectorizer"><tt class="xref py py-class docutils literal"><span class="pre">CountVectorizer</span></tt></a> and <a class="reference internal" href="generated/sklearn.feature_extraction.text.TfidfTransformer.html#sklearn.feature_extraction.text.TfidfTransformer" title="sklearn.feature_extraction.text.TfidfTransformer"><tt class="xref py py-class docutils literal"><span class="pre">TfidfTransformer</span></tt></a> in a single model:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">sklearn.feature_extraction.text</span> <span class="kn">import</span> <span class="n">TfidfVectorizer</span>
<span class="gp">>>> </span><span class="n">vectorizer</span> <span class="o">=</span> <span class="n">TfidfVectorizer</span><span class="p">(</span><span class="n">min_df</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">vectorizer</span><span class="o">.</span><span class="n">fit_transform</span><span class="p">(</span><span class="n">corpus</span><span class="p">)</span>
<span class="gp">... </span>
<span class="go"><4x9 sparse matrix of type '<... 'numpy.float64'>'</span>
<span class="go"> with 19 stored elements in Compressed Sparse ... format></span>
</pre></div>
</div>
<p>While the tf–idf normalization is often very useful, there might
be cases where the binary occurrence markers might offer better
features. This can be achieved by using the <tt class="docutils literal"><span class="pre">binary</span></tt> parameter
of <a class="reference internal" href="generated/sklearn.feature_extraction.text.CountVectorizer.html#sklearn.feature_extraction.text.CountVectorizer" title="sklearn.feature_extraction.text.CountVectorizer"><tt class="xref py py-class docutils literal"><span class="pre">CountVectorizer</span></tt></a>. In particular, some estimators such as
<a class="reference internal" href="naive_bayes.html#bernoulli-naive-bayes"><em>Bernoulli Naive Bayes</em></a> explicitly model discrete boolean random
variables. Also, very short texts are likely to have noisy tf–idf values
while the binary occurrence info is more stable.</p>
<p>As usual the best way to adjust the feature extraction parameters
is to use a cross-validated grid search, for instance by pipelining the
feature extractor with a classifier:</p>
<blockquote>
<div><ul class="simple">
<li><a class="reference internal" href="../auto_examples/model_selection/grid_search_text_feature_extraction.html#example-model-selection-grid-search-text-feature-extraction-py"><em>Sample pipeline for text feature extraction and evaluation</em></a></li>
</ul>
</div></blockquote>
</div>
<div class="section" id="decoding-text-files">
<h3>4.2.3.5. Decoding text files<a class="headerlink" href="#decoding-text-files" title="Permalink to this headline">¶</a></h3>
<p>Text is made of characters, but files are made of bytes. These bytes represent
characters according to some <em>encoding</em>. To work with text files in Python,
their bytes must be <em>decoded</em> to a character set called Unicode.
Common encodings are ASCII, Latin-1 (Western Europe), KOI8-R (Russian)
and the universal encodings UTF-8 and UTF-16. Many others exist.</p>
<div class="admonition note">
<p class="first admonition-title">Note</p>
<p class="last">An encoding can also be called a ‘character set’,
but this term is less accurate: several encodings can exist
for a single character set.</p>
</div>
<p>The text feature extractors in scikit-learn know how to decode text files,
but only if you tell them what encoding the files are in.
The <a class="reference internal" href="generated/sklearn.feature_extraction.text.CountVectorizer.html#sklearn.feature_extraction.text.CountVectorizer" title="sklearn.feature_extraction.text.CountVectorizer"><tt class="xref py py-class docutils literal"><span class="pre">CountVectorizer</span></tt></a> takes an <tt class="docutils literal"><span class="pre">encoding</span></tt> parameter for this purpose.
For modern text files, the correct encoding is probably UTF-8,
which is therefore the default (<tt class="docutils literal"><span class="pre">encoding="utf-8"</span></tt>).</p>
<p>If the text you are loading is not actually encoded with UTF-8, however,
you will get a <tt class="docutils literal"><span class="pre">UnicodeDecodeError</span></tt>.
The vectorizers can be told to be silent about decoding errors
by setting the <tt class="docutils literal"><span class="pre">decode_error</span></tt> parameter to either <tt class="docutils literal"><span class="pre">"ignore"</span></tt>
or <tt class="docutils literal"><span class="pre">"replace"</span></tt>. See the documentation for the Python function
<tt class="docutils literal"><span class="pre">bytes.decode</span></tt> for more details
(type <tt class="docutils literal"><span class="pre">help(bytes.decode)</span></tt> at the Python prompt).</p>
<p>If you are having trouble decoding text, here are some things to try:</p>
<ul class="simple">
<li>Find out what the actual encoding of the text is. The file might come
with a header or README that tells you the encoding, or there might be some
standard encoding you can assume based on where the text comes from.</li>
<li>You may be able to find out what kind of encoding it is in general
using the UNIX command <tt class="docutils literal"><span class="pre">file</span></tt>. The Python <tt class="docutils literal"><span class="pre">chardet</span></tt> module comes with
a script called <tt class="docutils literal"><span class="pre">chardetect.py</span></tt> that will guess the specific encoding,
though you cannot rely on its guess being correct.</li>
<li>You could try UTF-8 and disregard the errors. You can decode byte
strings with <tt class="docutils literal"><span class="pre">bytes.decode(errors='replace')</span></tt> to replace all
decoding errors with a meaningless character, or set
<tt class="docutils literal"><span class="pre">decode_error='replace'</span></tt> in the vectorizer. This may damage the
usefulness of your features.</li>
<li>Real text may come from a variety of sources that may have used different
encodings, or even be sloppily decoded in a different encoding than the
one it was encoded with. This is common in text retrieved from the Web.
The Python package <a class="reference external" href="http://github.com/LuminosoInsight/python-ftfy">ftfy</a> can automatically sort out some classes of
decoding errors, so you could try decoding the unknown text as <tt class="docutils literal"><span class="pre">latin-1</span></tt>
and then using <tt class="docutils literal"><span class="pre">ftfy</span></tt> to fix errors.</li>
<li>If the text is in a mish-mash of encodings that is simply too hard to sort
out (which is the case for the 20 Newsgroups dataset), you can fall back on
a simple single-byte encoding such as <tt class="docutils literal"><span class="pre">latin-1</span></tt>. Some text may display
incorrectly, but at least the same sequence of bytes will always represent
the same feature.</li>
</ul>
<p>For example, the following snippet uses <tt class="docutils literal"><span class="pre">chardet</span></tt>
(not shipped with scikit-learn, must be installed separately)
to figure out the encoding of three texts.
It then vectorizes the texts and prints the learned vocabulary.
The output is not shown here.</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="kn">import</span> <span class="nn">chardet</span>
<span class="gp">>>> </span><span class="n">text1</span> <span class="o">=</span> <span class="n">b</span><span class="s">"Sei mir gegr</span><span class="se">\xc3\xbc\xc3\x9f</span><span class="s">t mein Sauerkraut"</span>
<span class="gp">>>> </span><span class="n">text2</span> <span class="o">=</span> <span class="n">b</span><span class="s">"holdselig sind deine Ger</span><span class="se">\xfc</span><span class="s">che"</span>
<span class="gp">>>> </span><span class="n">text3</span> <span class="o">=</span> <span class="n">b</span><span class="s">"</span><span class="se">\xff\xfe</span><span class="s">A</span><span class="se">\x00</span><span class="s">u</span><span class="se">\x00</span><span class="s">f</span><span class="se">\x00</span><span class="s"> </span><span class="se">\x00</span><span class="s">F</span><span class="se">\x00</span><span class="s">l</span><span class="se">\x00\xfc\x00</span><span class="s">g</span><span class="se">\x00</span><span class="s">e</span><span class="se">\x00</span><span class="s">l</span><span class="se">\x00</span><span class="s">n</span><span class="se">\x00</span><span class="s"> </span><span class="se">\x00</span><span class="s">d</span><span class="se">\x00</span><span class="s">e</span><span class="se">\x00</span><span class="s">s</span><span class="se">\x00</span><span class="s"> </span><span class="se">\x00</span><span class="s">G</span><span class="se">\x00</span><span class="s">e</span><span class="se">\x00</span><span class="s">s</span><span class="se">\x00</span><span class="s">a</span><span class="se">\x00</span><span class="s">n</span><span class="se">\x00</span><span class="s">g</span><span class="se">\x00</span><span class="s">e</span><span class="se">\x00</span><span class="s">s</span><span class="se">\x00</span><span class="s">,</span><span class="se">\x00</span><span class="s"> </span><span class="se">\x00</span><span class="s">H</span><span class="se">\x00</span><span class="s">e</span><span class="se">\x00</span><span class="s">r</span><span class="se">\x00</span><span class="s">z</span><span class="se">\x00</span><span class="s">l</span><span class="se">\x00</span><span class="s">i</span><span class="se">\x00</span><span class="s">e</span><span class="se">\x00</span><span class="s">b</span><span class="se">\x00</span><span class="s">c</span><span class="se">\x00</span><span class="s">h</span><span class="se">\x00</span><span class="s">e</span><span class="se">\x00</span><span class="s">n</span><span class="se">\x00</span><span class="s">,</span><span class="se">\x00</span><span class="s"> </span><span class="se">\x00</span><span class="s">t</span><span class="se">\x00</span><span class="s">r</span><span class="se">\x00</span><span class="s">a</span><span class="se">\x00</span><span class="s">g</span><span class="se">\x00</span><span class="s"> </span><span class="se">\x00</span><span class="s">i</span><span class="se">\x00</span><span class="s">c</span><span class="se">\x00</span><span class="s">h</span><span class="se">\x00</span><span class="s"> </span><span class="se">\x00</span><span class="s">d</span><span class="se">\x00</span><span class="s">i</span><span class="se">\x00</span><span class="s">c</span><span class="se">\x00</span><span class="s">h</span><span class="se">\x00</span><span class="s"> </span><span class="se">\x00</span><span class="s">f</span><span class="se">\x00</span><span class="s">o</span><span class="se">\x00</span><span class="s">r</span><span class="se">\x00</span><span class="s">t</span><span class="se">\x00</span><span class="s">"</span>
<span class="gp">>>> </span><span class="n">decoded</span> <span class="o">=</span> <span class="p">[</span><span class="n">x</span><span class="o">.</span><span class="n">decode</span><span class="p">(</span><span class="n">chardet</span><span class="o">.</span><span class="n">detect</span><span class="p">(</span><span class="n">x</span><span class="p">)[</span><span class="s">'encoding'</span><span class="p">])</span>
<span class="gp">... </span> <span class="k">for</span> <span class="n">x</span> <span class="ow">in</span> <span class="p">(</span><span class="n">text1</span><span class="p">,</span> <span class="n">text2</span><span class="p">,</span> <span class="n">text3</span><span class="p">)]</span>
<span class="gp">>>> </span><span class="n">v</span> <span class="o">=</span> <span class="n">CountVectorizer</span><span class="p">()</span><span class="o">.</span><span class="n">fit</span><span class="p">(</span><span class="n">decoded</span><span class="p">)</span><span class="o">.</span><span class="n">vocabulary_</span>
<span class="gp">>>> </span><span class="k">for</span> <span class="n">term</span> <span class="ow">in</span> <span class="n">v</span><span class="p">:</span> <span class="k">print</span><span class="p">(</span><span class="n">v</span><span class="p">)</span>
</pre></div>
</div>
<p>(Depending on the version of <tt class="docutils literal"><span class="pre">chardet</span></tt>, it might get the first one wrong.)</p>
<p>For an introduction to Unicode and character encodings in general,
see Joel Spolsky’s <a class="reference external" href="http://www.joelonsoftware.com/articles/Unicode.html">Absolute Minimum Every Software Developer Must Know
About Unicode</a>.</p>
</div>
<div class="section" id="applications-and-examples">
<h3>4.2.3.6. Applications and examples<a class="headerlink" href="#applications-and-examples" title="Permalink to this headline">¶</a></h3>
<p>The bag of words representation is quite simplistic but surprisingly
useful in practice.</p>
<p>In particular in a <strong>supervised setting</strong> it can be successfully combined
with fast and scalable linear models to train <strong>document classifiers</strong>,
for instance:</p>
<blockquote>
<div><ul class="simple">
<li><a class="reference internal" href="../auto_examples/text/document_classification_20newsgroups.html#example-text-document-classification-20newsgroups-py"><em>Classification of text documents using sparse features</em></a></li>
</ul>
</div></blockquote>
<p>In an <strong>unsupervised setting</strong> it can be used to group similar documents
together by applying clustering algorithms such as <a class="reference internal" href="clustering.html#k-means"><em>K-means</em></a>:</p>
<blockquote>
<div><ul class="simple">
<li><a class="reference internal" href="../auto_examples/text/document_clustering.html#example-text-document-clustering-py"><em>Clustering text documents using k-means</em></a></li>
</ul>
</div></blockquote>
<p>Finally it is possible to discover the main topics of a corpus by
relaxing the hard assignment constraint of clustering, for instance by
using <a class="reference internal" href="decomposition.html#nmf"><em>Non-negative matrix factorization (NMF or NNMF)</em></a>:</p>
<blockquote>
<div><ul class="simple">
<li><a class="reference internal" href="../auto_examples/applications/topics_extraction_with_nmf.html#example-applications-topics-extraction-with-nmf-py"><em>Topics extraction with Non-Negative Matrix Factorization</em></a></li>
</ul>
</div></blockquote>
</div>
<div class="section" id="limitations-of-the-bag-of-words-representation">
<h3>4.2.3.7. Limitations of the Bag of Words representation<a class="headerlink" href="#limitations-of-the-bag-of-words-representation" title="Permalink to this headline">¶</a></h3>
<p>A collection of unigrams (what bag of words is) cannot capture phrases
and multi-word expressions, effectively disregarding any word order
dependence. Additionally, the bag of words model doesn’t account for potential
misspellings or word derivations.</p>
<p>N-grams to the rescue! Instead of building a simple collection of
unigrams (n=1), one might prefer a collection of bigrams (n=2), where
occurrences of pairs of consecutive words are counted.</p>
<p>One might alternatively consider a collection of character n-grams, a
representation resilient against misspellings and derivations.</p>
<p>For example, let’s say we’re dealing with a corpus of two documents:
<tt class="docutils literal"><span class="pre">['words',</span> <span class="pre">'wprds']</span></tt>. The second document contains a misspelling
of the word ‘words’.
A simple bag of words representation would consider these two as
very distinct documents, differing in both of the two possible features.
A character 2-gram representation, however, would find the documents
matching in 4 out of 8 features, which may help the preferred classifier
decide better:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="n">ngram_vectorizer</span> <span class="o">=</span> <span class="n">CountVectorizer</span><span class="p">(</span><span class="n">analyzer</span><span class="o">=</span><span class="s">'char_wb'</span><span class="p">,</span> <span class="n">ngram_range</span><span class="o">=</span><span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="mi">2</span><span class="p">),</span> <span class="n">min_df</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">counts</span> <span class="o">=</span> <span class="n">ngram_vectorizer</span><span class="o">.</span><span class="n">fit_transform</span><span class="p">([</span><span class="s">'words'</span><span class="p">,</span> <span class="s">'wprds'</span><span class="p">])</span>
<span class="gp">>>> </span><span class="n">ngram_vectorizer</span><span class="o">.</span><span class="n">get_feature_names</span><span class="p">()</span> <span class="o">==</span> <span class="p">(</span>
<span class="gp">... </span> <span class="p">[</span><span class="s">' w'</span><span class="p">,</span> <span class="s">'ds'</span><span class="p">,</span> <span class="s">'or'</span><span class="p">,</span> <span class="s">'pr'</span><span class="p">,</span> <span class="s">'rd'</span><span class="p">,</span> <span class="s">'s '</span><span class="p">,</span> <span class="s">'wo'</span><span class="p">,</span> <span class="s">'wp'</span><span class="p">])</span>
<span class="go">True</span>
<span class="gp">>>> </span><span class="n">counts</span><span class="o">.</span><span class="n">toarray</span><span class="p">()</span><span class="o">.</span><span class="n">astype</span><span class="p">(</span><span class="nb">int</span><span class="p">)</span>
<span class="go">array([[1, 1, 1, 0, 1, 1, 1, 0],</span>
<span class="go"> [1, 1, 0, 1, 1, 1, 0, 1]])</span>
</pre></div>
</div>
<p>In the above example, <tt class="docutils literal"><span class="pre">'char_wb</span></tt> analyzer is used, which creates n-grams
only from characters inside word boundaries (padded with space on each
side). The <tt class="docutils literal"><span class="pre">'char'</span></tt> analyzer, alternatively, creates n-grams that
span across words:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="n">ngram_vectorizer</span> <span class="o">=</span> <span class="n">CountVectorizer</span><span class="p">(</span><span class="n">analyzer</span><span class="o">=</span><span class="s">'char_wb'</span><span class="p">,</span> <span class="n">ngram_range</span><span class="o">=</span><span class="p">(</span><span class="mi">5</span><span class="p">,</span> <span class="mi">5</span><span class="p">),</span> <span class="n">min_df</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">ngram_vectorizer</span><span class="o">.</span><span class="n">fit_transform</span><span class="p">([</span><span class="s">'jumpy fox'</span><span class="p">])</span>
<span class="gp">... </span>
<span class="go"><1x4 sparse matrix of type '<... 'numpy.int64'>'</span>
<span class="go"> with 4 stored elements in Compressed Sparse ... format></span>
<span class="gp">>>> </span><span class="n">ngram_vectorizer</span><span class="o">.</span><span class="n">get_feature_names</span><span class="p">()</span> <span class="o">==</span> <span class="p">(</span>
<span class="gp">... </span> <span class="p">[</span><span class="s">' fox '</span><span class="p">,</span> <span class="s">' jump'</span><span class="p">,</span> <span class="s">'jumpy'</span><span class="p">,</span> <span class="s">'umpy '</span><span class="p">])</span>
<span class="go">True</span>
<span class="gp">>>> </span><span class="n">ngram_vectorizer</span> <span class="o">=</span> <span class="n">CountVectorizer</span><span class="p">(</span><span class="n">analyzer</span><span class="o">=</span><span class="s">'char'</span><span class="p">,</span> <span class="n">ngram_range</span><span class="o">=</span><span class="p">(</span><span class="mi">5</span><span class="p">,</span> <span class="mi">5</span><span class="p">),</span> <span class="n">min_df</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">ngram_vectorizer</span><span class="o">.</span><span class="n">fit_transform</span><span class="p">([</span><span class="s">'jumpy fox'</span><span class="p">])</span>
<span class="gp">... </span>
<span class="go"><1x5 sparse matrix of type '<... 'numpy.int64'>'</span>
<span class="go"> with 5 stored elements in Compressed Sparse ... format></span>
<span class="gp">>>> </span><span class="n">ngram_vectorizer</span><span class="o">.</span><span class="n">get_feature_names</span><span class="p">()</span> <span class="o">==</span> <span class="p">(</span>
<span class="gp">... </span> <span class="p">[</span><span class="s">'jumpy'</span><span class="p">,</span> <span class="s">'mpy f'</span><span class="p">,</span> <span class="s">'py fo'</span><span class="p">,</span> <span class="s">'umpy '</span><span class="p">,</span> <span class="s">'y fox'</span><span class="p">])</span>
<span class="go">True</span>
</pre></div>
</div>
<p>The word boundaries-aware variant <tt class="docutils literal"><span class="pre">char_wb</span></tt> is especially interesting
for languages that use white-spaces for word separation as it generates
significantly less noisy features than the raw <tt class="docutils literal"><span class="pre">char</span></tt> variant in
that case. For such languages it can increase both the predictive
accuracy and convergence speed of classifiers trained using such
features while retaining the robustness with regards to misspellings and
word derivations.</p>
<p>While some local positioning information can be preserved by extracting
n-grams instead of individual words, bag of words and bag of n-grams
destroy most of the inner structure of the document and hence most of
the meaning carried by that internal structure.</p>
<p>In order to address the wider task of Natural Language Understanding,
the local structure of sentences and paragraphs should thus be taken
into account. Many such models will thus be casted as “Structured output”
problems which are currently outside of the scope of scikit-learn.</p>
</div>
<div class="section" id="vectorizing-a-large-text-corpus-with-the-hashing-trick">
<span id="hashing-vectorizer"></span><h3>4.2.3.8. Vectorizing a large text corpus with the hashing trick<a class="headerlink" href="#vectorizing-a-large-text-corpus-with-the-hashing-trick" title="Permalink to this headline">¶</a></h3>
<p>The above vectorization scheme is simple but the fact that it holds an <strong>in-
memory mapping from the string tokens to the integer feature indices</strong> (the
<tt class="docutils literal"><span class="pre">vocabulary_</span></tt> attribute) causes several <strong>problems when dealing with large
datasets</strong>:</p>
<ul class="simple">
<li>the larger the corpus, the larger the vocabulary will grow and hence the
memory use too,</li>
<li>fitting requires the allocation of intermediate data structures
of size proportional to that of the original dataset.</li>
<li>building the word-mapping requires a full pass over the dataset hence it is
not possible to fit text classifiers in a strictly online manner.</li>
<li>pickling and un-pickling vectorizers with a large <tt class="docutils literal"><span class="pre">vocabulary_</span></tt> can be very
slow (typically much slower than pickling / un-pickling flat data structures
such as a NumPy array of the same size),</li>
<li>it is not easily possible to split the vectorization work into concurrent sub
tasks as the <tt class="docutils literal"><span class="pre">vocabulary_</span></tt> attribute would have to be a shared state with a
fine grained synchronization barrier: the mapping from token string to
feature index is dependent on ordering of the first occurrence of each token
hence would have to be shared, potentially harming the concurrent workers’
performance to the point of making them slower than the sequential variant.</li>
</ul>
<p>It is possible to overcome those limitations by combining the “hashing trick”
(<a class="reference internal" href="#feature-hashing"><em>Feature hashing</em></a>) implemented by the
<a class="reference internal" href="generated/sklearn.feature_extraction.FeatureHasher.html#sklearn.feature_extraction.FeatureHasher" title="sklearn.feature_extraction.FeatureHasher"><tt class="xref py py-class docutils literal"><span class="pre">sklearn.feature_extraction.FeatureHasher</span></tt></a> class and the text
preprocessing and tokenization features of the <a class="reference internal" href="generated/sklearn.feature_extraction.text.CountVectorizer.html#sklearn.feature_extraction.text.CountVectorizer" title="sklearn.feature_extraction.text.CountVectorizer"><tt class="xref py py-class docutils literal"><span class="pre">CountVectorizer</span></tt></a>.</p>
<p>This combination is implementing in <a class="reference internal" href="generated/sklearn.feature_extraction.text.HashingVectorizer.html#sklearn.feature_extraction.text.HashingVectorizer" title="sklearn.feature_extraction.text.HashingVectorizer"><tt class="xref py py-class docutils literal"><span class="pre">HashingVectorizer</span></tt></a>,
a transformer class that is mostly API compatible with <a class="reference internal" href="generated/sklearn.feature_extraction.text.CountVectorizer.html#sklearn.feature_extraction.text.CountVectorizer" title="sklearn.feature_extraction.text.CountVectorizer"><tt class="xref py py-class docutils literal"><span class="pre">CountVectorizer</span></tt></a>.
<a class="reference internal" href="generated/sklearn.feature_extraction.text.HashingVectorizer.html#sklearn.feature_extraction.text.HashingVectorizer" title="sklearn.feature_extraction.text.HashingVectorizer"><tt class="xref py py-class docutils literal"><span class="pre">HashingVectorizer</span></tt></a> is stateless,
meaning that you don’t have to call <tt class="docutils literal"><span class="pre">fit</span></tt> on it:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">sklearn.feature_extraction.text</span> <span class="kn">import</span> <span class="n">HashingVectorizer</span>
<span class="gp">>>> </span><span class="n">hv</span> <span class="o">=</span> <span class="n">HashingVectorizer</span><span class="p">(</span><span class="n">n_features</span><span class="o">=</span><span class="mi">10</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">hv</span><span class="o">.</span><span class="n">transform</span><span class="p">(</span><span class="n">corpus</span><span class="p">)</span>
<span class="gp">... </span>
<span class="go"><4x10 sparse matrix of type '<... 'numpy.float64'>'</span>
<span class="go"> with 16 stored elements in Compressed Sparse ... format></span>
</pre></div>
</div>
<p>You can see that 16 non-zero feature tokens were extracted in the vector
output: this is less than the 19 non-zeros extracted previously by the
<a class="reference internal" href="generated/sklearn.feature_extraction.text.CountVectorizer.html#sklearn.feature_extraction.text.CountVectorizer" title="sklearn.feature_extraction.text.CountVectorizer"><tt class="xref py py-class docutils literal"><span class="pre">CountVectorizer</span></tt></a> on the same toy corpus. The discrepancy comes from
hash function collisions because of the low value of the <tt class="docutils literal"><span class="pre">n_features</span></tt> parameter.</p>
<p>In a real world setting, the <tt class="docutils literal"><span class="pre">n_features</span></tt> parameter can be left to its
default value of <tt class="docutils literal"><span class="pre">2</span> <span class="pre">**</span> <span class="pre">20</span></tt> (roughly one million possible features). If memory
or downstream models size is an issue selecting a lower value such as <tt class="docutils literal"><span class="pre">2</span> <span class="pre">**</span>
<span class="pre">18</span></tt> might help without introducing too many additional collisions on typical
text classification tasks.</p>
<p>Note that the dimensionality does not affect the CPU training time of
algorithms which operate on CSR matrices (<tt class="docutils literal"><span class="pre">LinearSVC(dual=True)</span></tt>,
<tt class="docutils literal"><span class="pre">Perceptron</span></tt>, <tt class="docutils literal"><span class="pre">SGDClassifier</span></tt>, <tt class="docutils literal"><span class="pre">PassiveAggressive</span></tt>) but it does for
algorithms that work with CSC matrices (<tt class="docutils literal"><span class="pre">LinearSVC(dual=False)</span></tt>, <tt class="docutils literal"><span class="pre">Lasso()</span></tt>,
etc).</p>
<p>Let’s try again with the default setting:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="n">hv</span> <span class="o">=</span> <span class="n">HashingVectorizer</span><span class="p">()</span>
<span class="gp">>>> </span><span class="n">hv</span><span class="o">.</span><span class="n">transform</span><span class="p">(</span><span class="n">corpus</span><span class="p">)</span>
<span class="gp">... </span>
<span class="go"><4x1048576 sparse matrix of type '<... 'numpy.float64'>'</span>
<span class="go"> with 19 stored elements in Compressed Sparse ... format></span>
</pre></div>
</div>
<p>We no longer get the collisions, but this comes at the expense of a much larger
dimensionality of the output space.
Of course, other terms than the 19 used here
might still collide with each other.</p>
<p>The <a class="reference internal" href="generated/sklearn.feature_extraction.text.HashingVectorizer.html#sklearn.feature_extraction.text.HashingVectorizer" title="sklearn.feature_extraction.text.HashingVectorizer"><tt class="xref py py-class docutils literal"><span class="pre">HashingVectorizer</span></tt></a> also comes with the following limitations:</p>
<ul class="simple">
<li>it is not possible to invert the model (no <tt class="docutils literal"><span class="pre">inverse_transform</span></tt> method),
nor to access the original string representation of the features,
because of the one-way nature of the hash function that performs the mapping.</li>
<li>it does not provide IDF weighting as that would introduce statefulness in the
model. A <a class="reference internal" href="generated/sklearn.feature_extraction.text.TfidfTransformer.html#sklearn.feature_extraction.text.TfidfTransformer" title="sklearn.feature_extraction.text.TfidfTransformer"><tt class="xref py py-class docutils literal"><span class="pre">TfidfTransformer</span></tt></a> can be appended to it in a pipeline if
required.</li>
</ul>
</div>
<div class="section" id="performing-out-of-core-scaling-with-hashingvectorizer">
<h3>4.2.3.9. Performing out-of-core scaling with HashingVectorizer<a class="headerlink" href="#performing-out-of-core-scaling-with-hashingvectorizer" title="Permalink to this headline">¶</a></h3>
<p>An interesting development of using a <a class="reference internal" href="generated/sklearn.feature_extraction.text.HashingVectorizer.html#sklearn.feature_extraction.text.HashingVectorizer" title="sklearn.feature_extraction.text.HashingVectorizer"><tt class="xref py py-class docutils literal"><span class="pre">HashingVectorizer</span></tt></a> is the ability
to perform <a class="reference external" href="http://en.wikipedia.org/wiki/Out-of-core_algorithm">out-of-core</a> scaling. This means that we can learn from data that
does not fit into the computer’s main memory.</p>
<p>A strategy to implement out-of-core scaling is to stream data to the estimator
in mini-batches. Each mini-batch is vectorized using <a class="reference internal" href="generated/sklearn.feature_extraction.text.HashingVectorizer.html#sklearn.feature_extraction.text.HashingVectorizer" title="sklearn.feature_extraction.text.HashingVectorizer"><tt class="xref py py-class docutils literal"><span class="pre">HashingVectorizer</span></tt></a>
so as to guarantee that the input space of the estimator has always the same
dimensionality. The amount of memory used at any time is thus bounded by the
size of a mini-batch. Although there is no limit to the amount of data that can
be ingested using such an approach, from a practical point of view the learning
time is often limited by the CPU time one wants to spend on the task.</p>
<p>For a full-fledged example of out-of-core scaling in a text classification
task see <a class="reference internal" href="../auto_examples/applications/plot_out_of_core_classification.html#example-applications-plot-out-of-core-classification-py"><em>Out-of-core classification of text documents</em></a>.</p>
</div>
<div class="section" id="customizing-the-vectorizer-classes">
<h3>4.2.3.10. Customizing the vectorizer classes<a class="headerlink" href="#customizing-the-vectorizer-classes" title="Permalink to this headline">¶</a></h3>
<p>It is possible to customize the behavior by passing a callable
to the vectorizer constructor:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="k">def</span> <span class="nf">my_tokenizer</span><span class="p">(</span><span class="n">s</span><span class="p">):</span>
<span class="gp">... </span> <span class="k">return</span> <span class="n">s</span><span class="o">.</span><span class="n">split</span><span class="p">()</span>
<span class="gp">...</span>
<span class="gp">>>> </span><span class="n">vectorizer</span> <span class="o">=</span> <span class="n">CountVectorizer</span><span class="p">(</span><span class="n">tokenizer</span><span class="o">=</span><span class="n">my_tokenizer</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">vectorizer</span><span class="o">.</span><span class="n">build_analyzer</span><span class="p">()(</span><span class="s">u"Some... punctuation!"</span><span class="p">)</span> <span class="o">==</span> <span class="p">(</span>
<span class="gp">... </span> <span class="p">[</span><span class="s">'some...'</span><span class="p">,</span> <span class="s">'punctuation!'</span><span class="p">])</span>
<span class="go">True</span>
</pre></div>
</div>
<p>In particular we name:</p>
<blockquote>
<div><ul class="simple">
<li><tt class="docutils literal"><span class="pre">preprocessor</span></tt>: a callable that takes an entire document as input (as a
single string), and returns a possibly transformed version of the document,
still as an entire string. This can be used to remove HTML tags, lowercase
the entire document, etc.</li>
<li><tt class="docutils literal"><span class="pre">tokenizer</span></tt>: a callable that takes the output from the preprocessor
and splits it into tokens, then returns a list of these.</li>
<li><tt class="docutils literal"><span class="pre">analyzer</span></tt>: a callable that replaces the preprocessor and tokenizer.
The default analyzers all call the preprocessor and tokenizer, but custom
analyzers will skip this. N-gram extraction and stop word filtering take
place at the analyzer level, so a custom analyzer may have to reproduce
these steps.</li>
</ul>
</div></blockquote>
<p>(Lucene users might recognize these names, but be aware that scikit-learn
concepts may not map one-to-one onto Lucene concepts.)</p>
<p>To make the preprocessor, tokenizer and analyzers aware of the model
parameters it is possible to derive from the class and override the
<tt class="docutils literal"><span class="pre">build_preprocessor</span></tt>, <tt class="docutils literal"><span class="pre">build_tokenizer`</span></tt> and <tt class="docutils literal"><span class="pre">build_analyzer</span></tt>
factory methods instead of passing custom functions.</p>
<p>Some tips and tricks:</p>
<blockquote>
<div><ul>
<li><p class="first">If documents are pre-tokenized by an external package, then store them in
files (or strings) with the tokens separated by whitespace and pass
<tt class="docutils literal"><span class="pre">analyzer=str.split</span></tt></p>
</li>
<li><p class="first">Fancy token-level analysis such as stemming, lemmatizing, compound
splitting, filtering based on part-of-speech, etc. are not included in the
scikit-learn codebase, but can be added by customizing either the
tokenizer or the analyzer.
Here’s a <tt class="docutils literal"><span class="pre">CountVectorizer</span></tt> with a tokenizer and lemmatizer using
<a class="reference external" href="http://www.nltk.org">NLTK</a>:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">nltk</span> <span class="kn">import</span> <span class="n">word_tokenize</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">nltk.stem</span> <span class="kn">import</span> <span class="n">WordNetLemmatizer</span>
<span class="gp">>>> </span><span class="k">class</span> <span class="nc">LemmaTokenizer</span><span class="p">(</span><span class="nb">object</span><span class="p">):</span>
<span class="gp">... </span> <span class="k">def</span> <span class="nf">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="gp">... </span> <span class="bp">self</span><span class="o">.</span><span class="n">wnl</span> <span class="o">=</span> <span class="n">WordNetLemmatizer</span><span class="p">()</span>
<span class="gp">... </span> <span class="k">def</span> <span class="nf">__call__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">doc</span><span class="p">):</span>
<span class="gp">... </span> <span class="k">return</span> <span class="p">[</span><span class="bp">self</span><span class="o">.</span><span class="n">wnl</span><span class="o">.</span><span class="n">lemmatize</span><span class="p">(</span><span class="n">t</span><span class="p">)</span> <span class="k">for</span> <span class="n">t</span> <span class="ow">in</span> <span class="n">word_tokenize</span><span class="p">(</span><span class="n">doc</span><span class="p">)]</span>
<span class="gp">...</span>
<span class="gp">>>> </span><span class="n">vect</span> <span class="o">=</span> <span class="n">CountVectorizer</span><span class="p">(</span><span class="n">tokenizer</span><span class="o">=</span><span class="n">LemmaTokenizer</span><span class="p">())</span>
</pre></div>
</div>
<p>(Note that this will not filter out punctuation.)</p>
</li>
</ul>
</div></blockquote>
<p>Customizing the vectorizer can also be useful when handling Asian languages
that do not use an explicit word separator such as whitespace.</p>
</div>
</div>
<div class="section" id="image-feature-extraction">
<h2>4.2.4. Image feature extraction<a class="headerlink" href="#image-feature-extraction" title="Permalink to this headline">¶</a></h2>
<div class="section" id="patch-extraction">
<h3>4.2.4.1. Patch extraction<a class="headerlink" href="#patch-extraction" title="Permalink to this headline">¶</a></h3>
<p>The <a class="reference internal" href="generated/sklearn.feature_extraction.image.extract_patches_2d.html#sklearn.feature_extraction.image.extract_patches_2d" title="sklearn.feature_extraction.image.extract_patches_2d"><tt class="xref py py-func docutils literal"><span class="pre">extract_patches_2d</span></tt></a> function extracts patches from an image stored
as a two-dimensional array, or three-dimensional with color information along
the third axis. For rebuilding an image from all its patches, use
<a class="reference internal" href="generated/sklearn.feature_extraction.image.reconstruct_from_patches_2d.html#sklearn.feature_extraction.image.reconstruct_from_patches_2d" title="sklearn.feature_extraction.image.reconstruct_from_patches_2d"><tt class="xref py py-func docutils literal"><span class="pre">reconstruct_from_patches_2d</span></tt></a>. For example let use generate a 4x4 pixel
picture with 3 color channels (e.g. in RGB format):</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="kn">import</span> <span class="nn">numpy</span> <span class="kn">as</span> <span class="nn">np</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">sklearn.feature_extraction</span> <span class="kn">import</span> <span class="n">image</span>
<span class="gp">>>> </span><span class="n">one_image</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">arange</span><span class="p">(</span><span class="mi">4</span> <span class="o">*</span> <span class="mi">4</span> <span class="o">*</span> <span class="mi">3</span><span class="p">)</span><span class="o">.</span><span class="n">reshape</span><span class="p">((</span><span class="mi">4</span><span class="p">,</span> <span class="mi">4</span><span class="p">,</span> <span class="mi">3</span><span class="p">))</span>
<span class="gp">>>> </span><span class="n">one_image</span><span class="p">[:,</span> <span class="p">:,</span> <span class="mi">0</span><span class="p">]</span> <span class="c"># R channel of a fake RGB picture</span>
<span class="go">array([[ 0, 3, 6, 9],</span>
<span class="go"> [12, 15, 18, 21],</span>
<span class="go"> [24, 27, 30, 33],</span>
<span class="go"> [36, 39, 42, 45]])</span>
<span class="gp">>>> </span><span class="n">patches</span> <span class="o">=</span> <span class="n">image</span><span class="o">.</span><span class="n">extract_patches_2d</span><span class="p">(</span><span class="n">one_image</span><span class="p">,</span> <span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="mi">2</span><span class="p">),</span> <span class="n">max_patches</span><span class="o">=</span><span class="mi">2</span><span class="p">,</span>
<span class="gp">... </span> <span class="n">random_state</span><span class="o">=</span><span class="mi">0</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">patches</span><span class="o">.</span><span class="n">shape</span>
<span class="go">(2, 2, 2, 3)</span>
<span class="gp">>>> </span><span class="n">patches</span><span class="p">[:,</span> <span class="p">:,</span> <span class="p">:,</span> <span class="mi">0</span><span class="p">]</span>
<span class="go">array([[[ 0, 3],</span>
<span class="go"> [12, 15]],</span>
<span class="go"> [[15, 18],</span>
<span class="go"> [27, 30]]])</span>
<span class="gp">>>> </span><span class="n">patches</span> <span class="o">=</span> <span class="n">image</span><span class="o">.</span><span class="n">extract_patches_2d</span><span class="p">(</span><span class="n">one_image</span><span class="p">,</span> <span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="mi">2</span><span class="p">))</span>
<span class="gp">>>> </span><span class="n">patches</span><span class="o">.</span><span class="n">shape</span>
<span class="go">(9, 2, 2, 3)</span>
<span class="gp">>>> </span><span class="n">patches</span><span class="p">[</span><span class="mi">4</span><span class="p">,</span> <span class="p">:,</span> <span class="p">:,</span> <span class="mi">0</span><span class="p">]</span>
<span class="go">array([[15, 18],</span>
<span class="go"> [27, 30]])</span>
</pre></div>
</div>
<p>Let us now try to reconstruct the original image from the patches by averaging
on overlapping areas:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="n">reconstructed</span> <span class="o">=</span> <span class="n">image</span><span class="o">.</span><span class="n">reconstruct_from_patches_2d</span><span class="p">(</span><span class="n">patches</span><span class="p">,</span> <span class="p">(</span><span class="mi">4</span><span class="p">,</span> <span class="mi">4</span><span class="p">,</span> <span class="mi">3</span><span class="p">))</span>
<span class="gp">>>> </span><span class="n">np</span><span class="o">.</span><span class="n">testing</span><span class="o">.</span><span class="n">assert_array_equal</span><span class="p">(</span><span class="n">one_image</span><span class="p">,</span> <span class="n">reconstructed</span><span class="p">)</span>
</pre></div>
</div>
<p>The <a class="reference internal" href="generated/sklearn.feature_extraction.image.PatchExtractor.html#sklearn.feature_extraction.image.PatchExtractor" title="sklearn.feature_extraction.image.PatchExtractor"><tt class="xref py py-class docutils literal"><span class="pre">PatchExtractor</span></tt></a> class works in the same way as
<a class="reference internal" href="generated/sklearn.feature_extraction.image.extract_patches_2d.html#sklearn.feature_extraction.image.extract_patches_2d" title="sklearn.feature_extraction.image.extract_patches_2d"><tt class="xref py py-func docutils literal"><span class="pre">extract_patches_2d</span></tt></a>, only it supports multiple images as input. It is
implemented as an estimator, so it can be used in pipelines. See:</p>
<div class="highlight-python"><div class="highlight"><pre><span class="gp">>>> </span><span class="n">five_images</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">arange</span><span class="p">(</span><span class="mi">5</span> <span class="o">*</span> <span class="mi">4</span> <span class="o">*</span> <span class="mi">4</span> <span class="o">*</span> <span class="mi">3</span><span class="p">)</span><span class="o">.</span><span class="n">reshape</span><span class="p">(</span><span class="mi">5</span><span class="p">,</span> <span class="mi">4</span><span class="p">,</span> <span class="mi">4</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">patches</span> <span class="o">=</span> <span class="n">image</span><span class="o">.</span><span class="n">PatchExtractor</span><span class="p">((</span><span class="mi">2</span><span class="p">,</span> <span class="mi">2</span><span class="p">))</span><span class="o">.</span><span class="n">transform</span><span class="p">(</span><span class="n">five_images</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">patches</span><span class="o">.</span><span class="n">shape</span>
<span class="go">(45, 2, 2, 3)</span>
</pre></div>
</div>