forked from scikit-learn/scikit-learn.github.io
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathensemble.html
1901 lines (1844 loc) · 195 KB
/
ensemble.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>
<!--[if IE 8]><html class="no-js lt-ie9" lang="en" > <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="en" > <!--<![endif]-->
<head>
<meta charset="utf-8">
<meta name="generator" content="Docutils 0.19: https://docutils.sourceforge.io/" />
<meta property="og:title" content="1.11. Ensembles: Gradient boosting, random forests, bagging, voting, stacking" />
<meta property="og:type" content="website" />
<meta property="og:url" content="https://scikit-learn/stable/modules/ensemble.html" />
<meta property="og:site_name" content="scikit-learn" />
<meta property="og:description" content="Ensemble methods combine the predictions of several base estimators built with a given learning algorithm in order to improve generalizability / robustness over a single estimator. Two very famous ..." />
<meta property="og:image" content="https://scikit-learn/stable/_images/sphx_glr_plot_gradient_boosting_regression_001.png" />
<meta property="og:image:alt" content="scikit-learn" />
<meta name="description" content="Ensemble methods combine the predictions of several base estimators built with a given learning algorithm in order to improve generalizability / robustness over a single estimator. Two very famous ..." />
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>1.11. Ensembles: Gradient boosting, random forests, bagging, voting, stacking — scikit-learn 1.4.dev0 documentation</title>
<link rel="canonical" href="http://scikit-learn.org/stable/modules/ensemble.html" />
<link rel="shortcut icon" href="../_static/favicon.ico"/>
<link rel="stylesheet" href="../_static/css/vendor/bootstrap.min.css" type="text/css" />
<link rel="stylesheet" href="../_static/pygments.css" type="text/css" />
<link rel="stylesheet" href="../_static/css/theme.css" type="text/css" />
<link rel="stylesheet" href="../_static/copybutton.css" type="text/css" />
<link rel="stylesheet" href="../_static/plot_directive.css" type="text/css" />
<link rel="stylesheet" href="https://fonts.googleapis.com/css?family=Vibur" type="text/css" />
<link rel="stylesheet" href="../_static/jupyterlite_sphinx.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-binder.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-dataframe.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-rendered-html.css" type="text/css" />
<link rel="stylesheet" href="../_static/css/theme.css" type="text/css" />
<script id="documentation_options" data-url_root="../" src="../_static/documentation_options.js"></script>
<script src="../_static/js/vendor/jquery-3.6.3.slim.min.js"></script>
</head>
<body>
<nav id="navbar" class="sk-docs-navbar navbar navbar-expand-md navbar-light bg-light py-0">
<div class="container-fluid sk-docs-container px-0">
<a class="navbar-brand py-0" href="../index.html">
<img
class="sk-brand-img"
src="../_static/scikit-learn-logo-small.png"
alt="logo"/>
</a>
<button
id="sk-navbar-toggler"
class="navbar-toggler"
type="button"
data-toggle="collapse"
data-target="#navbarSupportedContent"
aria-controls="navbarSupportedContent"
aria-expanded="false"
aria-label="Toggle navigation"
>
<span class="navbar-toggler-icon"></span>
</button>
<div class="sk-navbar-collapse collapse navbar-collapse" id="navbarSupportedContent">
<ul class="navbar-nav mr-auto">
<li class="nav-item">
<a class="sk-nav-link nav-link" href="../install.html">Install</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link" href="../user_guide.html">User Guide</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link" href="classes.html">API</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link" href="../auto_examples/index.html">Examples</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link" target="_blank" rel="noopener noreferrer" href="https://blog.scikit-learn.org/">Community</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../getting_started.html" >Getting Started</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../tutorial/index.html" >Tutorial</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../whats_new/v1.4.html" >What's new</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../glossary.html" >Glossary</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../developers/index.html" >Development</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../faq.html" >FAQ</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../support.html" >Support</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../related_projects.html" >Related packages</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../roadmap.html" >Roadmap</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../governance.html" >Governance</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../about.html" >About us</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="https://github.com/scikit-learn/scikit-learn" >GitHub</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="https://scikit-learn.org/dev/versions.html" >Other Versions and Download</a>
</li>
<li class="nav-item dropdown nav-more-item-dropdown">
<a class="sk-nav-link nav-link dropdown-toggle" href="#" id="navbarDropdown" role="button" data-toggle="dropdown" aria-haspopup="true" aria-expanded="false">More</a>
<div class="dropdown-menu" aria-labelledby="navbarDropdown">
<a class="sk-nav-dropdown-item dropdown-item" href="../getting_started.html" >Getting Started</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../tutorial/index.html" >Tutorial</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../whats_new/v1.4.html" >What's new</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../glossary.html" >Glossary</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../developers/index.html" >Development</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../faq.html" >FAQ</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../support.html" >Support</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../related_projects.html" >Related packages</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../roadmap.html" >Roadmap</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../governance.html" >Governance</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../about.html" >About us</a>
<a class="sk-nav-dropdown-item dropdown-item" href="https://github.com/scikit-learn/scikit-learn" >GitHub</a>
<a class="sk-nav-dropdown-item dropdown-item" href="https://scikit-learn.org/dev/versions.html" >Other Versions and Download</a>
</div>
</li>
</ul>
<div id="searchbox" role="search">
<div class="searchformwrapper">
<form class="search" action="../search.html" method="get">
<input class="sk-search-text-input" type="text" name="q" aria-labelledby="searchlabel" />
<input class="sk-search-text-btn" type="submit" value="Go" />
</form>
</div>
</div>
</div>
</div>
</nav>
<div class="d-flex" id="sk-doc-wrapper">
<input type="checkbox" name="sk-toggle-checkbox" id="sk-toggle-checkbox">
<label id="sk-sidemenu-toggle" class="sk-btn-toggle-toc btn sk-btn-primary" for="sk-toggle-checkbox">Toggle Menu</label>
<div id="sk-sidebar-wrapper" class="border-right">
<div class="sk-sidebar-toc-wrapper">
<div class="btn-group w-100 mb-2" role="group" aria-label="rellinks">
<a href="tree.html" role="button" class="btn sk-btn-rellink py-1" sk-rellink-tooltip="1.10. Decision Trees">Prev</a><a href="../supervised_learning.html" role="button" class="btn sk-btn-rellink py-1" sk-rellink-tooltip="1. Supervised learning">Up</a>
<a href="multiclass.html" role="button" class="btn sk-btn-rellink py-1" sk-rellink-tooltip="1.12. Multiclass and multioutput algorithms">Next</a>
</div>
<div class="alert alert-danger p-1 mb-2" role="alert">
<p class="text-center mb-0">
<strong>scikit-learn 1.4.dev0</strong><br/>
<a href="http://scikit-learn.org/dev/versions.html">Other versions</a>
</p>
</div>
<div class="alert alert-warning p-1 mb-2" role="alert">
<p class="text-center mb-0">
Please <a class="font-weight-bold" href="../about.html#citing-scikit-learn"><string>cite us</string></a> if you use the software.
</p>
</div>
<div class="sk-sidebar-toc">
<ul>
<li><a class="reference internal" href="#">1.11. Ensembles: Gradient boosting, random forests, bagging, voting, stacking</a><ul>
<li><a class="reference internal" href="#gradient-boosting">1.11.1. Gradient-boosted trees</a><ul>
<li><a class="reference internal" href="#histogram-based-gradient-boosting">1.11.1.1. Histogram-Based Gradient Boosting</a><ul>
<li><a class="reference internal" href="#usage">1.11.1.1.1. Usage</a></li>
<li><a class="reference internal" href="#missing-values-support">1.11.1.1.2. Missing values support</a></li>
<li><a class="reference internal" href="#sample-weight-support">1.11.1.1.3. Sample weight support</a></li>
<li><a class="reference internal" href="#categorical-features-support">1.11.1.1.4. Categorical Features Support</a></li>
<li><a class="reference internal" href="#monotonic-constraints">1.11.1.1.5. Monotonic Constraints</a></li>
<li><a class="reference internal" href="#interaction-constraints">1.11.1.1.6. Interaction constraints</a></li>
<li><a class="reference internal" href="#low-level-parallelism">1.11.1.1.7. Low-level parallelism</a></li>
<li><a class="reference internal" href="#why-it-s-faster">1.11.1.1.8. Why it’s faster</a></li>
</ul>
</li>
<li><a class="reference internal" href="#gradientboostingclassifier-and-gradientboostingregressor">1.11.1.2. <code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingClassifier</span></code> and <code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingRegressor</span></code></a><ul>
<li><a class="reference internal" href="#classification">1.11.1.2.1. Classification</a></li>
<li><a class="reference internal" href="#regression">1.11.1.2.2. Regression</a></li>
<li><a class="reference internal" href="#fitting-additional-weak-learners">1.11.1.2.3. Fitting additional weak-learners</a></li>
<li><a class="reference internal" href="#controlling-the-tree-size">1.11.1.2.4. Controlling the tree size</a></li>
<li><a class="reference internal" href="#mathematical-formulation">1.11.1.2.5. Mathematical formulation</a><ul>
<li><a class="reference internal" href="#id10">1.11.1.2.5.1. Regression</a></li>
<li><a class="reference internal" href="#id11">1.11.1.2.5.2. Classification</a></li>
</ul>
</li>
<li><a class="reference internal" href="#loss-functions">1.11.1.2.6. Loss Functions</a></li>
<li><a class="reference internal" href="#shrinkage-via-learning-rate">1.11.1.2.7. Shrinkage via learning rate</a></li>
<li><a class="reference internal" href="#subsampling">1.11.1.2.8. Subsampling</a></li>
<li><a class="reference internal" href="#interpretation-with-feature-importance">1.11.1.2.9. Interpretation with feature importance</a></li>
</ul>
</li>
</ul>
</li>
<li><a class="reference internal" href="#random-forests-and-other-randomized-tree-ensembles">1.11.2. Random forests and other randomized tree ensembles</a><ul>
<li><a class="reference internal" href="#id18">1.11.2.1. Random Forests</a></li>
<li><a class="reference internal" href="#extremely-randomized-trees">1.11.2.2. Extremely Randomized Trees</a></li>
<li><a class="reference internal" href="#parameters">1.11.2.3. Parameters</a></li>
<li><a class="reference internal" href="#parallelization">1.11.2.4. Parallelization</a></li>
<li><a class="reference internal" href="#feature-importance-evaluation">1.11.2.5. Feature importance evaluation</a></li>
<li><a class="reference internal" href="#totally-random-trees-embedding">1.11.2.6. Totally Random Trees Embedding</a></li>
</ul>
</li>
<li><a class="reference internal" href="#bagging-meta-estimator">1.11.3. Bagging meta-estimator</a></li>
<li><a class="reference internal" href="#voting-classifier">1.11.4. Voting Classifier</a><ul>
<li><a class="reference internal" href="#majority-class-labels-majority-hard-voting">1.11.4.1. Majority Class Labels (Majority/Hard Voting)</a></li>
<li><a class="reference internal" href="#id26">1.11.4.2. Usage</a></li>
<li><a class="reference internal" href="#weighted-average-probabilities-soft-voting">1.11.4.3. Weighted Average Probabilities (Soft Voting)</a></li>
<li><a class="reference internal" href="#using-the-votingclassifier-with-gridsearchcv">1.11.4.4. Using the <code class="docutils literal notranslate"><span class="pre">VotingClassifier</span></code> with <code class="docutils literal notranslate"><span class="pre">GridSearchCV</span></code></a></li>
<li><a class="reference internal" href="#id27">1.11.4.5. Usage</a></li>
</ul>
</li>
<li><a class="reference internal" href="#voting-regressor">1.11.5. Voting Regressor</a><ul>
<li><a class="reference internal" href="#id29">1.11.5.1. Usage</a></li>
</ul>
</li>
<li><a class="reference internal" href="#stacked-generalization">1.11.6. Stacked generalization</a></li>
<li><a class="reference internal" href="#adaboost">1.11.7. AdaBoost</a><ul>
<li><a class="reference internal" href="#id37">1.11.7.1. Usage</a></li>
</ul>
</li>
</ul>
</li>
</ul>
</div>
</div>
</div>
<div id="sk-page-content-wrapper">
<div class="sk-page-content container-fluid body px-md-3" role="main">
<section id="ensembles-gradient-boosting-random-forests-bagging-voting-stacking">
<span id="ensemble"></span><h1><span class="section-number">1.11. </span>Ensembles: Gradient boosting, random forests, bagging, voting, stacking<a class="headerlink" href="#ensembles-gradient-boosting-random-forests-bagging-voting-stacking" title="Permalink to this heading">¶</a></h1>
<p><strong>Ensemble methods</strong> combine the predictions of several
base estimators built with a given learning algorithm in order to improve
generalizability / robustness over a single estimator.</p>
<p>Two very famous examples of ensemble methods are <a class="reference external" href="gradient_boosting">gradient-boosted trees</a> and <a class="reference external" href="forest">random forests</a>.</p>
<p>More generally, ensemble models can be applied to any base learner beyond
trees, in averaging methods such as <a class="reference internal" href="#bagging"><span class="std std-ref">Bagging methods</span></a>,
<a class="reference external" href="stacking">model stacking</a>, or <a class="reference external" href="voting_classifier">Voting</a>, or in
boosting, as <a class="reference internal" href="#adaboost"><span class="std std-ref">AdaBoost</span></a>.</p>
<nav class="contents local" id="contents">
<ul class="simple">
<li><p><a class="reference internal" href="#gradient-boosting" id="id38">Gradient-boosted trees</a></p></li>
<li><p><a class="reference internal" href="#random-forests-and-other-randomized-tree-ensembles" id="id39">Random forests and other randomized tree ensembles</a></p></li>
<li><p><a class="reference internal" href="#bagging-meta-estimator" id="id40">Bagging meta-estimator</a></p></li>
<li><p><a class="reference internal" href="#voting-classifier" id="id41">Voting Classifier</a></p></li>
<li><p><a class="reference internal" href="#voting-regressor" id="id42">Voting Regressor</a></p></li>
<li><p><a class="reference internal" href="#stacked-generalization" id="id43">Stacked generalization</a></p></li>
<li><p><a class="reference internal" href="#adaboost" id="id44">AdaBoost</a></p></li>
</ul>
</nav>
<section id="gradient-boosting">
<span id="id1"></span><h2><a class="toc-backref" href="#id38" role="doc-backlink"><span class="section-number">1.11.1. </span>Gradient-boosted trees</a><a class="headerlink" href="#gradient-boosting" title="Permalink to this heading">¶</a></h2>
<p><a class="reference external" href="https://en.wikipedia.org/wiki/Gradient_boosting">Gradient Tree Boosting</a>
or Gradient Boosted Decision Trees (GBDT) is a generalization
of boosting to arbitrary differentiable loss functions, see the seminal work of
<a class="reference internal" href="#friedman2001" id="id2"><span>[Friedman2001]</span></a>. GBDT is an excellent model for both regression and
classification, in particular for tabular data.</p>
<aside class="topic">
<p class="topic-title"><a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingClassifier.html#sklearn.ensemble.GradientBoostingClassifier" title="sklearn.ensemble.GradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingClassifier</span></code></a> vs <a class="reference internal" href="generated/sklearn.ensemble.HistGradientBoostingClassifier.html#sklearn.ensemble.HistGradientBoostingClassifier" title="sklearn.ensemble.HistGradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">HistGradientBoostingClassifier</span></code></a></p>
<p>Scikit-learn provides two implementations of gradient-boosted trees:
<a class="reference internal" href="generated/sklearn.ensemble.HistGradientBoostingClassifier.html#sklearn.ensemble.HistGradientBoostingClassifier" title="sklearn.ensemble.HistGradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">HistGradientBoostingClassifier</span></code></a> vs
<a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingClassifier.html#sklearn.ensemble.GradientBoostingClassifier" title="sklearn.ensemble.GradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingClassifier</span></code></a> for classification, and the
corresponding classes for regression. The former can be <strong>orders of
magnitude faster</strong> than the latter when the number of samples is
larger than tens of thousands of samples.</p>
<p>Missing values and categorical data are natively supported by the
Hist… version, removing the need for additional preprocessing such as
imputation.</p>
<p><a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingClassifier.html#sklearn.ensemble.GradientBoostingClassifier" title="sklearn.ensemble.GradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingClassifier</span></code></a> and
<a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingRegressor.html#sklearn.ensemble.GradientBoostingRegressor" title="sklearn.ensemble.GradientBoostingRegressor"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingRegressor</span></code></a>, might be preferred for small sample
sizes since binning may lead to split points that are too approximate
in this setting.</p>
</aside>
<section id="histogram-based-gradient-boosting">
<span id="id3"></span><h3><span class="section-number">1.11.1.1. </span>Histogram-Based Gradient Boosting<a class="headerlink" href="#histogram-based-gradient-boosting" title="Permalink to this heading">¶</a></h3>
<p>Scikit-learn 0.21 introduced two new implementations of
gradient boosted trees, namely <a class="reference internal" href="generated/sklearn.ensemble.HistGradientBoostingClassifier.html#sklearn.ensemble.HistGradientBoostingClassifier" title="sklearn.ensemble.HistGradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">HistGradientBoostingClassifier</span></code></a>
and <a class="reference internal" href="generated/sklearn.ensemble.HistGradientBoostingRegressor.html#sklearn.ensemble.HistGradientBoostingRegressor" title="sklearn.ensemble.HistGradientBoostingRegressor"><code class="xref py py-class docutils literal notranslate"><span class="pre">HistGradientBoostingRegressor</span></code></a>, inspired by
<a class="reference external" href="https://github.com/Microsoft/LightGBM">LightGBM</a> (See <a class="reference internal" href="#lightgbm" id="id4"><span>[LightGBM]</span></a>).</p>
<p>These histogram-based estimators can be <strong>orders of magnitude faster</strong>
than <a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingClassifier.html#sklearn.ensemble.GradientBoostingClassifier" title="sklearn.ensemble.GradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingClassifier</span></code></a> and
<a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingRegressor.html#sklearn.ensemble.GradientBoostingRegressor" title="sklearn.ensemble.GradientBoostingRegressor"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingRegressor</span></code></a> when the number of samples is larger
than tens of thousands of samples.</p>
<p>They also have built-in support for missing values, which avoids the need
for an imputer.</p>
<p>These fast estimators first bin the input samples <code class="docutils literal notranslate"><span class="pre">X</span></code> into
integer-valued bins (typically 256 bins) which tremendously reduces the
number of splitting points to consider, and allows the algorithm to
leverage integer-based data structures (histograms) instead of relying on
sorted continuous values when building the trees. The API of these
estimators is slightly different, and some of the features from
<a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingClassifier.html#sklearn.ensemble.GradientBoostingClassifier" title="sklearn.ensemble.GradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingClassifier</span></code></a> and <a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingRegressor.html#sklearn.ensemble.GradientBoostingRegressor" title="sklearn.ensemble.GradientBoostingRegressor"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingRegressor</span></code></a>
are not yet supported, for instance some loss functions.</p>
<aside class="topic">
<p class="topic-title">Examples:</p>
<ul class="simple">
<li><p><a class="reference internal" href="../auto_examples/inspection/plot_partial_dependence.html#sphx-glr-auto-examples-inspection-plot-partial-dependence-py"><span class="std std-ref">Partial Dependence and Individual Conditional Expectation Plots</span></a></p></li>
</ul>
</aside>
<section id="usage">
<h4><span class="section-number">1.11.1.1.1. </span>Usage<a class="headerlink" href="#usage" title="Permalink to this heading">¶</a></h4>
<p>Most of the parameters are unchanged from
<a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingClassifier.html#sklearn.ensemble.GradientBoostingClassifier" title="sklearn.ensemble.GradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingClassifier</span></code></a> and <a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingRegressor.html#sklearn.ensemble.GradientBoostingRegressor" title="sklearn.ensemble.GradientBoostingRegressor"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingRegressor</span></code></a>.
One exception is the <code class="docutils literal notranslate"><span class="pre">max_iter</span></code> parameter that replaces <code class="docutils literal notranslate"><span class="pre">n_estimators</span></code>, and
controls the number of iterations of the boosting process:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">sklearn.ensemble</span> <span class="kn">import</span> <span class="n">HistGradientBoostingClassifier</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">sklearn.datasets</span> <span class="kn">import</span> <span class="n">make_hastie_10_2</span>
<span class="gp">>>> </span><span class="n">X</span><span class="p">,</span> <span class="n">y</span> <span class="o">=</span> <span class="n">make_hastie_10_2</span><span class="p">(</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">X_train</span><span class="p">,</span> <span class="n">X_test</span> <span class="o">=</span> <span class="n">X</span><span class="p">[:</span><span class="mi">2000</span><span class="p">],</span> <span class="n">X</span><span class="p">[</span><span class="mi">2000</span><span class="p">:]</span>
<span class="gp">>>> </span><span class="n">y_train</span><span class="p">,</span> <span class="n">y_test</span> <span class="o">=</span> <span class="n">y</span><span class="p">[:</span><span class="mi">2000</span><span class="p">],</span> <span class="n">y</span><span class="p">[</span><span class="mi">2000</span><span class="p">:]</span>
<span class="gp">>>> </span><span class="n">clf</span> <span class="o">=</span> <span class="n">HistGradientBoostingClassifier</span><span class="p">(</span><span class="n">max_iter</span><span class="o">=</span><span class="mi">100</span><span class="p">)</span><span class="o">.</span><span class="n">fit</span><span class="p">(</span><span class="n">X_train</span><span class="p">,</span> <span class="n">y_train</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">clf</span><span class="o">.</span><span class="n">score</span><span class="p">(</span><span class="n">X_test</span><span class="p">,</span> <span class="n">y_test</span><span class="p">)</span>
<span class="go">0.8965</span>
</pre></div>
</div>
<p>Available losses for regression are ‘squared_error’,
‘absolute_error’, which is less sensitive to outliers, and
‘poisson’, which is well suited to model counts and frequencies. For
classification, ‘log_loss’ is the only option. For binary classification it uses the
binary log loss, also known as binomial deviance or binary cross-entropy. For
<code class="docutils literal notranslate"><span class="pre">n_classes</span> <span class="pre">>=</span> <span class="pre">3</span></code>, it uses the multi-class log loss function, with multinomial deviance
and categorical cross-entropy as alternative names. The appropriate loss version is
selected based on <a class="reference internal" href="../glossary.html#term-y"><span class="xref std std-term">y</span></a> passed to <a class="reference internal" href="../glossary.html#term-fit"><span class="xref std std-term">fit</span></a>.</p>
<p>The size of the trees can be controlled through the <code class="docutils literal notranslate"><span class="pre">max_leaf_nodes</span></code>,
<code class="docutils literal notranslate"><span class="pre">max_depth</span></code>, and <code class="docutils literal notranslate"><span class="pre">min_samples_leaf</span></code> parameters.</p>
<p>The number of bins used to bin the data is controlled with the <code class="docutils literal notranslate"><span class="pre">max_bins</span></code>
parameter. Using less bins acts as a form of regularization. It is
generally recommended to use as many bins as possible (256), which is the default.</p>
<p>The <code class="docutils literal notranslate"><span class="pre">l2_regularization</span></code> parameter is a regularizer on the loss function and
corresponds to <span class="math notranslate nohighlight">\(\lambda\)</span> in equation (2) of <a class="reference internal" href="#xgboost" id="id5"><span>[XGBoost]</span></a>.</p>
<p>Note that <strong>early-stopping is enabled by default if the number of samples is
larger than 10,000</strong>. The early-stopping behaviour is controlled via the
<code class="docutils literal notranslate"><span class="pre">early_stopping</span></code>, <code class="docutils literal notranslate"><span class="pre">scoring</span></code>, <code class="docutils literal notranslate"><span class="pre">validation_fraction</span></code>,
<code class="docutils literal notranslate"><span class="pre">n_iter_no_change</span></code>, and <code class="docutils literal notranslate"><span class="pre">tol</span></code> parameters. It is possible to early-stop
using an arbitrary <a class="reference internal" href="../glossary.html#term-scorer"><span class="xref std std-term">scorer</span></a>, or just the training or validation loss.
Note that for technical reasons, using a scorer is significantly slower than
using the loss. By default, early-stopping is performed if there are at least
10,000 samples in the training set, and uses the validation loss.</p>
</section>
<section id="missing-values-support">
<h4><span class="section-number">1.11.1.1.2. </span>Missing values support<a class="headerlink" href="#missing-values-support" title="Permalink to this heading">¶</a></h4>
<p><a class="reference internal" href="generated/sklearn.ensemble.HistGradientBoostingClassifier.html#sklearn.ensemble.HistGradientBoostingClassifier" title="sklearn.ensemble.HistGradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">HistGradientBoostingClassifier</span></code></a> and
<a class="reference internal" href="generated/sklearn.ensemble.HistGradientBoostingRegressor.html#sklearn.ensemble.HistGradientBoostingRegressor" title="sklearn.ensemble.HistGradientBoostingRegressor"><code class="xref py py-class docutils literal notranslate"><span class="pre">HistGradientBoostingRegressor</span></code></a> have built-in support for missing
values (NaNs).</p>
<p>During training, the tree grower learns at each split point whether samples
with missing values should go to the left or right child, based on the
potential gain. When predicting, samples with missing values are assigned to
the left or right child consequently:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">sklearn.ensemble</span> <span class="kn">import</span> <span class="n">HistGradientBoostingClassifier</span>
<span class="gp">>>> </span><span class="kn">import</span> <span class="nn">numpy</span> <span class="k">as</span> <span class="nn">np</span>
<span class="gp">>>> </span><span class="n">X</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">array</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="mi">2</span><span class="p">,</span> <span class="n">np</span><span class="o">.</span><span class="n">nan</span><span class="p">])</span><span class="o">.</span><span class="n">reshape</span><span class="p">(</span><span class="o">-</span><span class="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">y</span> <span class="o">=</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="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">]</span>
<span class="gp">>>> </span><span class="n">gbdt</span> <span class="o">=</span> <span class="n">HistGradientBoostingClassifier</span><span class="p">(</span><span class="n">min_samples_leaf</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">fit</span><span class="p">(</span><span class="n">X</span><span class="p">,</span> <span class="n">y</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">gbdt</span><span class="o">.</span><span class="n">predict</span><span class="p">(</span><span class="n">X</span><span class="p">)</span>
<span class="go">array([0, 0, 1, 1])</span>
</pre></div>
</div>
<p>When the missingness pattern is predictive, the splits can be performed on
whether the feature value is missing or not:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">X</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">array</span><span class="p">([</span><span class="mi">0</span><span class="p">,</span> <span class="n">np</span><span class="o">.</span><span class="n">nan</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="n">np</span><span class="o">.</span><span class="n">nan</span><span class="p">])</span><span class="o">.</span><span class="n">reshape</span><span class="p">(</span><span class="o">-</span><span class="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">y</span> <span class="o">=</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="mi">0</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="n">gbdt</span> <span class="o">=</span> <span class="n">HistGradientBoostingClassifier</span><span class="p">(</span><span class="n">min_samples_leaf</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span>
<span class="gp">... </span> <span class="n">max_depth</span><span class="o">=</span><span class="mi">2</span><span class="p">,</span>
<span class="gp">... </span> <span class="n">learning_rate</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span>
<span class="gp">... </span> <span class="n">max_iter</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">fit</span><span class="p">(</span><span class="n">X</span><span class="p">,</span> <span class="n">y</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">gbdt</span><span class="o">.</span><span class="n">predict</span><span class="p">(</span><span class="n">X</span><span class="p">)</span>
<span class="go">array([0, 1, 0, 0, 1])</span>
</pre></div>
</div>
<p>If no missing values were encountered for a given feature during training,
then samples with missing values are mapped to whichever child has the most
samples.</p>
</section>
<section id="sample-weight-support">
<span id="sw-hgbdt"></span><h4><span class="section-number">1.11.1.1.3. </span>Sample weight support<a class="headerlink" href="#sample-weight-support" title="Permalink to this heading">¶</a></h4>
<p><a class="reference internal" href="generated/sklearn.ensemble.HistGradientBoostingClassifier.html#sklearn.ensemble.HistGradientBoostingClassifier" title="sklearn.ensemble.HistGradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">HistGradientBoostingClassifier</span></code></a> and
<a class="reference internal" href="generated/sklearn.ensemble.HistGradientBoostingRegressor.html#sklearn.ensemble.HistGradientBoostingRegressor" title="sklearn.ensemble.HistGradientBoostingRegressor"><code class="xref py py-class docutils literal notranslate"><span class="pre">HistGradientBoostingRegressor</span></code></a> support sample weights during
<a class="reference internal" href="../glossary.html#term-fit"><span class="xref std std-term">fit</span></a>.</p>
<p>The following toy example demonstrates that samples with a sample weight of zero are ignored:</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">X</span> <span class="o">=</span> <span class="p">[[</span><span class="mi">1</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">1</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">1</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">0</span><span class="p">,</span> <span class="mi">1</span><span class="p">]]</span>
<span class="gp">>>> </span><span class="n">y</span> <span class="o">=</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="mi">1</span><span class="p">,</span> <span class="mi">0</span><span class="p">]</span>
<span class="gp">>>> </span><span class="c1"># ignore the first 2 training samples by setting their weight to 0</span>
<span class="gp">>>> </span><span class="n">sample_weight</span> <span class="o">=</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="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">]</span>
<span class="gp">>>> </span><span class="n">gb</span> <span class="o">=</span> <span class="n">HistGradientBoostingClassifier</span><span class="p">(</span><span class="n">min_samples_leaf</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">gb</span><span class="o">.</span><span class="n">fit</span><span class="p">(</span><span class="n">X</span><span class="p">,</span> <span class="n">y</span><span class="p">,</span> <span class="n">sample_weight</span><span class="o">=</span><span class="n">sample_weight</span><span class="p">)</span>
<span class="go">HistGradientBoostingClassifier(...)</span>
<span class="gp">>>> </span><span class="n">gb</span><span class="o">.</span><span class="n">predict</span><span class="p">([[</span><span class="mi">1</span><span class="p">,</span> <span class="mi">0</span><span class="p">]])</span>
<span class="go">array([1])</span>
<span class="gp">>>> </span><span class="n">gb</span><span class="o">.</span><span class="n">predict_proba</span><span class="p">([[</span><span class="mi">1</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="mi">1</span><span class="p">]</span>
<span class="go">0.99...</span>
</pre></div>
</div>
<p>As you can see, the <code class="docutils literal notranslate"><span class="pre">[1,</span> <span class="pre">0]</span></code> is comfortably classified as <code class="docutils literal notranslate"><span class="pre">1</span></code> since the first
two samples are ignored due to their sample weights.</p>
<p>Implementation detail: taking sample weights into account amounts to
multiplying the gradients (and the hessians) by the sample weights. Note that
the binning stage (specifically the quantiles computation) does not take the
weights into account.</p>
</section>
<section id="categorical-features-support">
<span id="categorical-support-gbdt"></span><h4><span class="section-number">1.11.1.1.4. </span>Categorical Features Support<a class="headerlink" href="#categorical-features-support" title="Permalink to this heading">¶</a></h4>
<p><a class="reference internal" href="generated/sklearn.ensemble.HistGradientBoostingClassifier.html#sklearn.ensemble.HistGradientBoostingClassifier" title="sklearn.ensemble.HistGradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">HistGradientBoostingClassifier</span></code></a> and
<a class="reference internal" href="generated/sklearn.ensemble.HistGradientBoostingRegressor.html#sklearn.ensemble.HistGradientBoostingRegressor" title="sklearn.ensemble.HistGradientBoostingRegressor"><code class="xref py py-class docutils literal notranslate"><span class="pre">HistGradientBoostingRegressor</span></code></a> have native support for categorical
features: they can consider splits on non-ordered, categorical data.</p>
<p>For datasets with categorical features, using the native categorical support
is often better than relying on one-hot encoding
(<a class="reference internal" href="generated/sklearn.preprocessing.OneHotEncoder.html#sklearn.preprocessing.OneHotEncoder" title="sklearn.preprocessing.OneHotEncoder"><code class="xref py py-class docutils literal notranslate"><span class="pre">OneHotEncoder</span></code></a>), because one-hot encoding
requires more tree depth to achieve equivalent splits. It is also usually
better to rely on the native categorical support rather than to treat
categorical features as continuous (ordinal), which happens for ordinal-encoded
categorical data, since categories are nominal quantities where order does not
matter.</p>
<p>To enable categorical support, a boolean mask can be passed to the
<code class="docutils literal notranslate"><span class="pre">categorical_features</span></code> parameter, indicating which feature is categorical. In
the following, the first feature will be treated as categorical and the
second feature as numerical:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">gbdt</span> <span class="o">=</span> <span class="n">HistGradientBoostingClassifier</span><span class="p">(</span><span class="n">categorical_features</span><span class="o">=</span><span class="p">[</span><span class="kc">True</span><span class="p">,</span> <span class="kc">False</span><span class="p">])</span>
</pre></div>
</div>
<p>Equivalently, one can pass a list of integers indicating the indices of the
categorical features:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">gbdt</span> <span class="o">=</span> <span class="n">HistGradientBoostingClassifier</span><span class="p">(</span><span class="n">categorical_features</span><span class="o">=</span><span class="p">[</span><span class="mi">0</span><span class="p">])</span>
</pre></div>
</div>
<p>The cardinality of each categorical feature must be less than the <code class="docutils literal notranslate"><span class="pre">max_bins</span></code>
parameter, and each categorical feature is expected to be encoded in
<code class="docutils literal notranslate"><span class="pre">[0,</span> <span class="pre">max_bins</span> <span class="pre">-</span> <span class="pre">1]</span></code>. To that end, it might be useful to pre-process the data
with an <a class="reference internal" href="generated/sklearn.preprocessing.OrdinalEncoder.html#sklearn.preprocessing.OrdinalEncoder" title="sklearn.preprocessing.OrdinalEncoder"><code class="xref py py-class docutils literal notranslate"><span class="pre">OrdinalEncoder</span></code></a> as done in
<a class="reference internal" href="../auto_examples/ensemble/plot_gradient_boosting_categorical.html#sphx-glr-auto-examples-ensemble-plot-gradient-boosting-categorical-py"><span class="std std-ref">Categorical Feature Support in Gradient Boosting</span></a>.</p>
<p>If there are missing values during training, the missing values will be
treated as a proper category. If there are no missing values during training,
then at prediction time, missing values are mapped to the child node that has
the most samples (just like for continuous features). When predicting,
categories that were not seen during fit time will be treated as missing
values.</p>
<p><strong>Split finding with categorical features</strong>: The canonical way of considering
categorical splits in a tree is to consider
all of the <span class="math notranslate nohighlight">\(2^{K - 1} - 1\)</span> partitions, where <span class="math notranslate nohighlight">\(K\)</span> is the number of
categories. This can quickly become prohibitive when <span class="math notranslate nohighlight">\(K\)</span> is large.
Fortunately, since gradient boosting trees are always regression trees (even
for classification problems), there exist a faster strategy that can yield
equivalent splits. First, the categories of a feature are sorted according to
the variance of the target, for each category <code class="docutils literal notranslate"><span class="pre">k</span></code>. Once the categories are
sorted, one can consider <em>continuous partitions</em>, i.e. treat the categories
as if they were ordered continuous values (see Fisher <a class="reference internal" href="#fisher1958" id="id6"><span>[Fisher1958]</span></a> for a
formal proof). As a result, only <span class="math notranslate nohighlight">\(K - 1\)</span> splits need to be considered
instead of <span class="math notranslate nohighlight">\(2^{K - 1} - 1\)</span>. The initial sorting is a
<span class="math notranslate nohighlight">\(\mathcal{O}(K \log(K))\)</span> operation, leading to a total complexity of
<span class="math notranslate nohighlight">\(\mathcal{O}(K \log(K) + K)\)</span>, instead of <span class="math notranslate nohighlight">\(\mathcal{O}(2^K)\)</span>.</p>
<aside class="topic">
<p class="topic-title">Examples:</p>
<ul class="simple">
<li><p><a class="reference internal" href="../auto_examples/ensemble/plot_gradient_boosting_categorical.html#sphx-glr-auto-examples-ensemble-plot-gradient-boosting-categorical-py"><span class="std std-ref">Categorical Feature Support in Gradient Boosting</span></a></p></li>
</ul>
</aside>
</section>
<section id="monotonic-constraints">
<span id="monotonic-cst-gbdt"></span><h4><span class="section-number">1.11.1.1.5. </span>Monotonic Constraints<a class="headerlink" href="#monotonic-constraints" title="Permalink to this heading">¶</a></h4>
<p>Depending on the problem at hand, you may have prior knowledge indicating
that a given feature should in general have a positive (or negative) effect
on the target value. For example, all else being equal, a higher credit
score should increase the probability of getting approved for a loan.
Monotonic constraints allow you to incorporate such prior knowledge into the
model.</p>
<p>For a predictor <span class="math notranslate nohighlight">\(F\)</span> with two features:</p>
<blockquote>
<div><ul>
<li><dl>
<dt>a <strong>monotonic increase constraint</strong> is a constraint of the form:</dt><dd><div class="math notranslate nohighlight">
\[x_1 \leq x_1' \implies F(x_1, x_2) \leq F(x_1', x_2)\]</div>
</dd>
</dl>
</li>
<li><dl>
<dt>a <strong>monotonic decrease constraint</strong> is a constraint of the form:</dt><dd><div class="math notranslate nohighlight">
\[x_1 \leq x_1' \implies F(x_1, x_2) \geq F(x_1', x_2)\]</div>
</dd>
</dl>
</li>
</ul>
</div></blockquote>
<p>You can specify a monotonic constraint on each feature using the
<code class="docutils literal notranslate"><span class="pre">monotonic_cst</span></code> parameter. For each feature, a value of 0 indicates no
constraint, while 1 and -1 indicate a monotonic increase and
monotonic decrease constraint, respectively:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">sklearn.ensemble</span> <span class="kn">import</span> <span class="n">HistGradientBoostingRegressor</span>
<span class="go">... # monotonic increase, monotonic decrease, and no constraint on the 3 features</span>
<span class="gp">>>> </span><span class="n">gbdt</span> <span class="o">=</span> <span class="n">HistGradientBoostingRegressor</span><span class="p">(</span><span class="n">monotonic_cst</span><span class="o">=</span><span class="p">[</span><span class="mi">1</span><span class="p">,</span> <span class="o">-</span><span class="mi">1</span><span class="p">,</span> <span class="mi">0</span><span class="p">])</span>
</pre></div>
</div>
<p>In a binary classification context, imposing a monotonic increase (decrease) constraint means that higher values of the feature are supposed
to have a positive (negative) effect on the probability of samples
to belong to the positive class.</p>
<p>Nevertheless, monotonic constraints only marginally constrain feature effects on the output.
For instance, monotonic increase and decrease constraints cannot be used to enforce the
following modelling constraint:</p>
<blockquote>
<div><div class="math notranslate nohighlight">
\[x_1 \leq x_1' \implies F(x_1, x_2) \leq F(x_1', x_2')\]</div>
</div></blockquote>
<p>Also, monotonic constraints are not supported for multiclass classification.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>Since categories are unordered quantities, it is not possible to enforce
monotonic constraints on categorical features.</p>
</div>
<aside class="topic">
<p class="topic-title">Examples:</p>
<ul class="simple">
<li><p><a class="reference internal" href="../auto_examples/ensemble/plot_monotonic_constraints.html#sphx-glr-auto-examples-ensemble-plot-monotonic-constraints-py"><span class="std std-ref">Monotonic Constraints</span></a></p></li>
</ul>
</aside>
</section>
<section id="interaction-constraints">
<span id="interaction-cst-hgbt"></span><h4><span class="section-number">1.11.1.1.6. </span>Interaction constraints<a class="headerlink" href="#interaction-constraints" title="Permalink to this heading">¶</a></h4>
<p>A priori, the histogram gradient boosted trees are allowed to use any feature
to split a node into child nodes. This creates so called interactions between
features, i.e. usage of different features as split along a branch. Sometimes,
one wants to restrict the possible interactions, see <a class="reference internal" href="#mayer2022" id="id7"><span>[Mayer2022]</span></a>. This can be
done by the parameter <code class="docutils literal notranslate"><span class="pre">interaction_cst</span></code>, where one can specify the indices
of features that are allowed to interact.
For instance, with 3 features in total, <code class="docutils literal notranslate"><span class="pre">interaction_cst=[{0},</span> <span class="pre">{1},</span> <span class="pre">{2}]</span></code>
forbids all interactions.
The constraints <code class="docutils literal notranslate"><span class="pre">[{0,</span> <span class="pre">1},</span> <span class="pre">{1,</span> <span class="pre">2}]</span></code> specifies two groups of possibly
interacting features. Features 0 and 1 may interact with each other, as well
as features 1 and 2. But note that features 0 and 2 are forbidden to interact.
The following depicts a tree and the possible splits of the tree:</p>
<div class="highlight-none notranslate"><div class="highlight"><pre><span></span> 1 <- Both constraint groups could be applied from now on
/ \
1 2 <- Left split still fulfills both constraint groups.
/ \ / \ Right split at feature 2 has only group {1, 2} from now on.
</pre></div>
</div>
<p>LightGBM uses the same logic for overlapping groups.</p>
<p>Note that features not listed in <code class="docutils literal notranslate"><span class="pre">interaction_cst</span></code> are automatically
assigned an interaction group for themselves. With again 3 features, this
means that <code class="docutils literal notranslate"><span class="pre">[{0}]</span></code> is equivalent to <code class="docutils literal notranslate"><span class="pre">[{0},</span> <span class="pre">{1,</span> <span class="pre">2}]</span></code>.</p>
<aside class="topic">
<p class="topic-title">References</p>
<div role="list" class="citation-list">
<div class="citation" id="mayer2022" role="doc-biblioentry">
<span class="label"><span class="fn-bracket">[</span><a role="doc-backlink" href="#id7">Mayer2022</a><span class="fn-bracket">]</span></span>
<p>M. Mayer, S.C. Bourassa, M. Hoesli, and D.F. Scognamiglio.
2022. <a class="reference external" href="https://doi.org/10.3390/jrfm15050193">Machine Learning Applications to Land and Structure Valuation</a>.
Journal of Risk and Financial Management 15, no. 5: 193</p>
</div>
</div>
</aside>
</section>
<section id="low-level-parallelism">
<h4><span class="section-number">1.11.1.1.7. </span>Low-level parallelism<a class="headerlink" href="#low-level-parallelism" title="Permalink to this heading">¶</a></h4>
<p><a class="reference internal" href="generated/sklearn.ensemble.HistGradientBoostingClassifier.html#sklearn.ensemble.HistGradientBoostingClassifier" title="sklearn.ensemble.HistGradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">HistGradientBoostingClassifier</span></code></a> and
<a class="reference internal" href="generated/sklearn.ensemble.HistGradientBoostingRegressor.html#sklearn.ensemble.HistGradientBoostingRegressor" title="sklearn.ensemble.HistGradientBoostingRegressor"><code class="xref py py-class docutils literal notranslate"><span class="pre">HistGradientBoostingRegressor</span></code></a> use OpenMP
for parallelization through Cython. For more details on how to control the
number of threads, please refer to our <a class="reference internal" href="../computing/parallelism.html#parallelism"><span class="std std-ref">Parallelism</span></a> notes.</p>
<p>The following parts are parallelized:</p>
<ul class="simple">
<li><p>mapping samples from real values to integer-valued bins (finding the bin
thresholds is however sequential)</p></li>
<li><p>building histograms is parallelized over features</p></li>
<li><p>finding the best split point at a node is parallelized over features</p></li>
<li><p>during fit, mapping samples into the left and right children is
parallelized over samples</p></li>
<li><p>gradient and hessians computations are parallelized over samples</p></li>
<li><p>predicting is parallelized over samples</p></li>
</ul>
</section>
<section id="why-it-s-faster">
<span id="id8"></span><h4><span class="section-number">1.11.1.1.8. </span>Why it’s faster<a class="headerlink" href="#why-it-s-faster" title="Permalink to this heading">¶</a></h4>
<p>The bottleneck of a gradient boosting procedure is building the decision
trees. Building a traditional decision tree (as in the other GBDTs
<a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingClassifier.html#sklearn.ensemble.GradientBoostingClassifier" title="sklearn.ensemble.GradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingClassifier</span></code></a> and <a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingRegressor.html#sklearn.ensemble.GradientBoostingRegressor" title="sklearn.ensemble.GradientBoostingRegressor"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingRegressor</span></code></a>)
requires sorting the samples at each node (for
each feature). Sorting is needed so that the potential gain of a split point
can be computed efficiently. Splitting a single node has thus a complexity
of <span class="math notranslate nohighlight">\(\mathcal{O}(n_\text{features} \times n \log(n))\)</span> where <span class="math notranslate nohighlight">\(n\)</span>
is the number of samples at the node.</p>
<p><a class="reference internal" href="generated/sklearn.ensemble.HistGradientBoostingClassifier.html#sklearn.ensemble.HistGradientBoostingClassifier" title="sklearn.ensemble.HistGradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">HistGradientBoostingClassifier</span></code></a> and
<a class="reference internal" href="generated/sklearn.ensemble.HistGradientBoostingRegressor.html#sklearn.ensemble.HistGradientBoostingRegressor" title="sklearn.ensemble.HistGradientBoostingRegressor"><code class="xref py py-class docutils literal notranslate"><span class="pre">HistGradientBoostingRegressor</span></code></a>, in contrast, do not require sorting the
feature values and instead use a data-structure called a histogram, where the
samples are implicitly ordered. Building a histogram has a
<span class="math notranslate nohighlight">\(\mathcal{O}(n)\)</span> complexity, so the node splitting procedure has a
<span class="math notranslate nohighlight">\(\mathcal{O}(n_\text{features} \times n)\)</span> complexity, much smaller
than the previous one. In addition, instead of considering <span class="math notranslate nohighlight">\(n\)</span> split
points, we consider only <code class="docutils literal notranslate"><span class="pre">max_bins</span></code> split points, which might be much
smaller.</p>
<p>In order to build histograms, the input data <code class="docutils literal notranslate"><span class="pre">X</span></code> needs to be binned into
integer-valued bins. This binning procedure does require sorting the feature
values, but it only happens once at the very beginning of the boosting process
(not at each node, like in <a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingClassifier.html#sklearn.ensemble.GradientBoostingClassifier" title="sklearn.ensemble.GradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingClassifier</span></code></a> and
<a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingRegressor.html#sklearn.ensemble.GradientBoostingRegressor" title="sklearn.ensemble.GradientBoostingRegressor"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingRegressor</span></code></a>).</p>
<p>Finally, many parts of the implementation of
<a class="reference internal" href="generated/sklearn.ensemble.HistGradientBoostingClassifier.html#sklearn.ensemble.HistGradientBoostingClassifier" title="sklearn.ensemble.HistGradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">HistGradientBoostingClassifier</span></code></a> and
<a class="reference internal" href="generated/sklearn.ensemble.HistGradientBoostingRegressor.html#sklearn.ensemble.HistGradientBoostingRegressor" title="sklearn.ensemble.HistGradientBoostingRegressor"><code class="xref py py-class docutils literal notranslate"><span class="pre">HistGradientBoostingRegressor</span></code></a> are parallelized.</p>
<aside class="topic">
<p class="topic-title">References</p>
<div role="list" class="citation-list">
<div class="citation" id="xgboost" role="doc-biblioentry">
<span class="label"><span class="fn-bracket">[</span><a role="doc-backlink" href="#id5">XGBoost</a><span class="fn-bracket">]</span></span>
<p>Tianqi Chen, Carlos Guestrin, <a class="reference external" href="https://arxiv.org/abs/1603.02754">“XGBoost: A Scalable Tree
Boosting System”</a></p>
</div>
<div class="citation" id="lightgbm" role="doc-biblioentry">
<span class="label"><span class="fn-bracket">[</span><a role="doc-backlink" href="#id4">LightGBM</a><span class="fn-bracket">]</span></span>
<p>Ke et. al. <a class="reference external" href="https://papers.nips.cc/paper/6907-lightgbm-a-highly-efficient-gradient-boosting-decision-tree">“LightGBM: A Highly Efficient Gradient
BoostingDecision Tree”</a></p>
</div>
<div class="citation" id="fisher1958" role="doc-biblioentry">
<span class="label"><span class="fn-bracket">[</span><a role="doc-backlink" href="#id6">Fisher1958</a><span class="fn-bracket">]</span></span>
<p>Fisher, W.D. (1958). <a class="reference external" href="http://csiss.ncgia.ucsb.edu/SPACE/workshops/2004/SAC/files/fisher.pdf">“On Grouping for Maximum Homogeneity”</a>
Journal of the American Statistical Association, 53, 789-798.</p>
</div>
</div>
</aside>
</section>
</section>
<section id="gradientboostingclassifier-and-gradientboostingregressor">
<h3><span class="section-number">1.11.1.2. </span><a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingClassifier.html#sklearn.ensemble.GradientBoostingClassifier" title="sklearn.ensemble.GradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingClassifier</span></code></a> and <a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingRegressor.html#sklearn.ensemble.GradientBoostingRegressor" title="sklearn.ensemble.GradientBoostingRegressor"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingRegressor</span></code></a><a class="headerlink" href="#gradientboostingclassifier-and-gradientboostingregressor" title="Permalink to this heading">¶</a></h3>
<p>The usage and the parameters of <a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingClassifier.html#sklearn.ensemble.GradientBoostingClassifier" title="sklearn.ensemble.GradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingClassifier</span></code></a> and
<a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingRegressor.html#sklearn.ensemble.GradientBoostingRegressor" title="sklearn.ensemble.GradientBoostingRegressor"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingRegressor</span></code></a> are described below. The 2 most important
parameters of these estimators are <code class="docutils literal notranslate"><span class="pre">n_estimators</span></code> and <code class="docutils literal notranslate"><span class="pre">learning_rate</span></code>.</p>
<section id="classification">
<h4><span class="section-number">1.11.1.2.1. </span>Classification<a class="headerlink" href="#classification" title="Permalink to this heading">¶</a></h4>
<p><a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingClassifier.html#sklearn.ensemble.GradientBoostingClassifier" title="sklearn.ensemble.GradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingClassifier</span></code></a> supports both binary and multi-class
classification.
The following example shows how to fit a gradient boosting classifier
with 100 decision stumps as weak learners:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">sklearn.datasets</span> <span class="kn">import</span> <span class="n">make_hastie_10_2</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">sklearn.ensemble</span> <span class="kn">import</span> <span class="n">GradientBoostingClassifier</span>
<span class="gp">>>> </span><span class="n">X</span><span class="p">,</span> <span class="n">y</span> <span class="o">=</span> <span class="n">make_hastie_10_2</span><span class="p">(</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">X_train</span><span class="p">,</span> <span class="n">X_test</span> <span class="o">=</span> <span class="n">X</span><span class="p">[:</span><span class="mi">2000</span><span class="p">],</span> <span class="n">X</span><span class="p">[</span><span class="mi">2000</span><span class="p">:]</span>
<span class="gp">>>> </span><span class="n">y_train</span><span class="p">,</span> <span class="n">y_test</span> <span class="o">=</span> <span class="n">y</span><span class="p">[:</span><span class="mi">2000</span><span class="p">],</span> <span class="n">y</span><span class="p">[</span><span class="mi">2000</span><span class="p">:]</span>
<span class="gp">>>> </span><span class="n">clf</span> <span class="o">=</span> <span class="n">GradientBoostingClassifier</span><span class="p">(</span><span class="n">n_estimators</span><span class="o">=</span><span class="mi">100</span><span class="p">,</span> <span class="n">learning_rate</span><span class="o">=</span><span class="mf">1.0</span><span class="p">,</span>
<span class="gp">... </span> <span class="n">max_depth</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span> <span class="n">random_state</span><span class="o">=</span><span class="mi">0</span><span class="p">)</span><span class="o">.</span><span class="n">fit</span><span class="p">(</span><span class="n">X_train</span><span class="p">,</span> <span class="n">y_train</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">clf</span><span class="o">.</span><span class="n">score</span><span class="p">(</span><span class="n">X_test</span><span class="p">,</span> <span class="n">y_test</span><span class="p">)</span>
<span class="go">0.913...</span>
</pre></div>
</div>
<p>The number of weak learners (i.e. regression trees) is controlled by the
parameter <code class="docutils literal notranslate"><span class="pre">n_estimators</span></code>; <a class="reference internal" href="#gradient-boosting-tree-size"><span class="std std-ref">The size of each tree</span></a> can be controlled either by setting the tree
depth via <code class="docutils literal notranslate"><span class="pre">max_depth</span></code> or by setting the number of leaf nodes via
<code class="docutils literal notranslate"><span class="pre">max_leaf_nodes</span></code>. The <code class="docutils literal notranslate"><span class="pre">learning_rate</span></code> is a hyper-parameter in the range
(0.0, 1.0] that controls overfitting via <a class="reference internal" href="#gradient-boosting-shrinkage"><span class="std std-ref">shrinkage</span></a> .</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>Classification with more than 2 classes requires the induction
of <code class="docutils literal notranslate"><span class="pre">n_classes</span></code> regression trees at each iteration,
thus, the total number of induced trees equals
<code class="docutils literal notranslate"><span class="pre">n_classes</span> <span class="pre">*</span> <span class="pre">n_estimators</span></code>. For datasets with a large number
of classes we strongly recommend to use
<a class="reference internal" href="generated/sklearn.ensemble.HistGradientBoostingClassifier.html#sklearn.ensemble.HistGradientBoostingClassifier" title="sklearn.ensemble.HistGradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">HistGradientBoostingClassifier</span></code></a> as an alternative to
<a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingClassifier.html#sklearn.ensemble.GradientBoostingClassifier" title="sklearn.ensemble.GradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingClassifier</span></code></a> .</p>
</div>
</section>
<section id="regression">
<h4><span class="section-number">1.11.1.2.2. </span>Regression<a class="headerlink" href="#regression" title="Permalink to this heading">¶</a></h4>
<p><a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingRegressor.html#sklearn.ensemble.GradientBoostingRegressor" title="sklearn.ensemble.GradientBoostingRegressor"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingRegressor</span></code></a> supports a number of
<a class="reference internal" href="#gradient-boosting-loss"><span class="std std-ref">different loss functions</span></a>
for regression which can be specified via the argument
<code class="docutils literal notranslate"><span class="pre">loss</span></code>; the default loss function for regression is squared error
(<code class="docutils literal notranslate"><span class="pre">'squared_error'</span></code>).</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">import</span> <span class="nn">numpy</span> <span class="k">as</span> <span class="nn">np</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">sklearn.metrics</span> <span class="kn">import</span> <span class="n">mean_squared_error</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">sklearn.datasets</span> <span class="kn">import</span> <span class="n">make_friedman1</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">sklearn.ensemble</span> <span class="kn">import</span> <span class="n">GradientBoostingRegressor</span>
<span class="gp">>>> </span><span class="n">X</span><span class="p">,</span> <span class="n">y</span> <span class="o">=</span> <span class="n">make_friedman1</span><span class="p">(</span><span class="n">n_samples</span><span class="o">=</span><span class="mi">1200</span><span class="p">,</span> <span class="n">random_state</span><span class="o">=</span><span class="mi">0</span><span class="p">,</span> <span class="n">noise</span><span class="o">=</span><span class="mf">1.0</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">X_train</span><span class="p">,</span> <span class="n">X_test</span> <span class="o">=</span> <span class="n">X</span><span class="p">[:</span><span class="mi">200</span><span class="p">],</span> <span class="n">X</span><span class="p">[</span><span class="mi">200</span><span class="p">:]</span>
<span class="gp">>>> </span><span class="n">y_train</span><span class="p">,</span> <span class="n">y_test</span> <span class="o">=</span> <span class="n">y</span><span class="p">[:</span><span class="mi">200</span><span class="p">],</span> <span class="n">y</span><span class="p">[</span><span class="mi">200</span><span class="p">:]</span>
<span class="gp">>>> </span><span class="n">est</span> <span class="o">=</span> <span class="n">GradientBoostingRegressor</span><span class="p">(</span>
<span class="gp">... </span> <span class="n">n_estimators</span><span class="o">=</span><span class="mi">100</span><span class="p">,</span> <span class="n">learning_rate</span><span class="o">=</span><span class="mf">0.1</span><span class="p">,</span> <span class="n">max_depth</span><span class="o">=</span><span class="mi">1</span><span class="p">,</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">loss</span><span class="o">=</span><span class="s1">'squared_error'</span>
<span class="gp">... </span><span class="p">)</span><span class="o">.</span><span class="n">fit</span><span class="p">(</span><span class="n">X_train</span><span class="p">,</span> <span class="n">y_train</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">mean_squared_error</span><span class="p">(</span><span class="n">y_test</span><span class="p">,</span> <span class="n">est</span><span class="o">.</span><span class="n">predict</span><span class="p">(</span><span class="n">X_test</span><span class="p">))</span>
<span class="go">5.00...</span>
</pre></div>
</div>
<p>The figure below shows the results of applying <a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingRegressor.html#sklearn.ensemble.GradientBoostingRegressor" title="sklearn.ensemble.GradientBoostingRegressor"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingRegressor</span></code></a>
with least squares loss and 500 base learners to the diabetes dataset
(<a class="reference internal" href="generated/sklearn.datasets.load_diabetes.html#sklearn.datasets.load_diabetes" title="sklearn.datasets.load_diabetes"><code class="xref py py-func docutils literal notranslate"><span class="pre">sklearn.datasets.load_diabetes</span></code></a>).
The plot shows the train and test error at each iteration.
The train error at each iteration is stored in the
<code class="docutils literal notranslate"><span class="pre">train_score_</span></code> attribute of the gradient boosting model.
The test error at each iterations can be obtained
via the <a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingRegressor.html#sklearn.ensemble.GradientBoostingRegressor.staged_predict" title="sklearn.ensemble.GradientBoostingRegressor.staged_predict"><code class="xref py py-meth docutils literal notranslate"><span class="pre">staged_predict</span></code></a> method which returns a
generator that yields the predictions at each stage. Plots like these can be used
to determine the optimal number of trees (i.e. <code class="docutils literal notranslate"><span class="pre">n_estimators</span></code>) by early stopping.</p>
<figure class="align-center">
<a class="reference external image-reference" href="../auto_examples/ensemble/plot_gradient_boosting_regression.html"><img alt="../_images/sphx_glr_plot_gradient_boosting_regression_001.png" src="../_images/sphx_glr_plot_gradient_boosting_regression_001.png" style="width: 450.0px; height: 450.0px;" /></a>
</figure>
<aside class="topic">
<p class="topic-title">Examples:</p>
<ul class="simple">
<li><p><a class="reference internal" href="../auto_examples/ensemble/plot_gradient_boosting_regression.html#sphx-glr-auto-examples-ensemble-plot-gradient-boosting-regression-py"><span class="std std-ref">Gradient Boosting regression</span></a></p></li>
<li><p><a class="reference internal" href="../auto_examples/ensemble/plot_gradient_boosting_oob.html#sphx-glr-auto-examples-ensemble-plot-gradient-boosting-oob-py"><span class="std std-ref">Gradient Boosting Out-of-Bag estimates</span></a></p></li>
</ul>
</aside>
</section>
<section id="fitting-additional-weak-learners">
<span id="gradient-boosting-warm-start"></span><h4><span class="section-number">1.11.1.2.3. </span>Fitting additional weak-learners<a class="headerlink" href="#fitting-additional-weak-learners" title="Permalink to this heading">¶</a></h4>
<p>Both <a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingRegressor.html#sklearn.ensemble.GradientBoostingRegressor" title="sklearn.ensemble.GradientBoostingRegressor"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingRegressor</span></code></a> and <a class="reference internal" href="generated/sklearn.ensemble.GradientBoostingClassifier.html#sklearn.ensemble.GradientBoostingClassifier" title="sklearn.ensemble.GradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradientBoostingClassifier</span></code></a>
support <code class="docutils literal notranslate"><span class="pre">warm_start=True</span></code> which allows you to add more estimators to an already
fitted model.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">_</span> <span class="o">=</span> <span class="n">est</span><span class="o">.</span><span class="n">set_params</span><span class="p">(</span><span class="n">n_estimators</span><span class="o">=</span><span class="mi">200</span><span class="p">,</span> <span class="n">warm_start</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span> <span class="c1"># set warm_start and new nr of trees</span>
<span class="gp">>>> </span><span class="n">_</span> <span class="o">=</span> <span class="n">est</span><span class="o">.</span><span class="n">fit</span><span class="p">(</span><span class="n">X_train</span><span class="p">,</span> <span class="n">y_train</span><span class="p">)</span> <span class="c1"># fit additional 100 trees to est</span>
<span class="gp">>>> </span><span class="n">mean_squared_error</span><span class="p">(</span><span class="n">y_test</span><span class="p">,</span> <span class="n">est</span><span class="o">.</span><span class="n">predict</span><span class="p">(</span><span class="n">X_test</span><span class="p">))</span>
<span class="go">3.84...</span>
</pre></div>
</div>
</section>
<section id="controlling-the-tree-size">
<span id="gradient-boosting-tree-size"></span><h4><span class="section-number">1.11.1.2.4. </span>Controlling the tree size<a class="headerlink" href="#controlling-the-tree-size" title="Permalink to this heading">¶</a></h4>
<p>The size of the regression tree base learners defines the level of variable
interactions that can be captured by the gradient boosting model. In general,
a tree of depth <code class="docutils literal notranslate"><span class="pre">h</span></code> can capture interactions of order <code class="docutils literal notranslate"><span class="pre">h</span></code> .
There are two ways in which the size of the individual regression trees can
be controlled.</p>
<p>If you specify <code class="docutils literal notranslate"><span class="pre">max_depth=h</span></code> then complete binary trees
of depth <code class="docutils literal notranslate"><span class="pre">h</span></code> will be grown. Such trees will have (at most) <code class="docutils literal notranslate"><span class="pre">2**h</span></code> leaf nodes
and <code class="docutils literal notranslate"><span class="pre">2**h</span> <span class="pre">-</span> <span class="pre">1</span></code> split nodes.</p>
<p>Alternatively, you can control the tree size by specifying the number of
leaf nodes via the parameter <code class="docutils literal notranslate"><span class="pre">max_leaf_nodes</span></code>. In this case,
trees will be grown using best-first search where nodes with the highest improvement
in impurity will be expanded first.
A tree with <code class="docutils literal notranslate"><span class="pre">max_leaf_nodes=k</span></code> has <code class="docutils literal notranslate"><span class="pre">k</span> <span class="pre">-</span> <span class="pre">1</span></code> split nodes and thus can
model interactions of up to order <code class="docutils literal notranslate"><span class="pre">max_leaf_nodes</span> <span class="pre">-</span> <span class="pre">1</span></code> .</p>
<p>We found that <code class="docutils literal notranslate"><span class="pre">max_leaf_nodes=k</span></code> gives comparable results to <code class="docutils literal notranslate"><span class="pre">max_depth=k-1</span></code>
but is significantly faster to train at the expense of a slightly higher
training error.
The parameter <code class="docutils literal notranslate"><span class="pre">max_leaf_nodes</span></code> corresponds to the variable <code class="docutils literal notranslate"><span class="pre">J</span></code> in the
chapter on gradient boosting in <a class="reference internal" href="#friedman2001" id="id9"><span>[Friedman2001]</span></a> and is related to the parameter
<code class="docutils literal notranslate"><span class="pre">interaction.depth</span></code> in R’s gbm package where <code class="docutils literal notranslate"><span class="pre">max_leaf_nodes</span> <span class="pre">==</span> <span class="pre">interaction.depth</span> <span class="pre">+</span> <span class="pre">1</span></code> .</p>
</section>
<section id="mathematical-formulation">
<h4><span class="section-number">1.11.1.2.5. </span>Mathematical formulation<a class="headerlink" href="#mathematical-formulation" title="Permalink to this heading">¶</a></h4>
<p>We first present GBRT for regression, and then detail the classification
case.</p>
<section id="id10">
<h5><span class="section-number">1.11.1.2.5.1. </span>Regression<a class="headerlink" href="#id10" title="Permalink to this heading">¶</a></h5>
<p>GBRT regressors are additive models whose prediction <span class="math notranslate nohighlight">\(\hat{y}_i\)</span> for a
given input <span class="math notranslate nohighlight">\(x_i\)</span> is of the following form:</p>
<blockquote>
<div><div class="math notranslate nohighlight">
\[\hat{y}_i = F_M(x_i) = \sum_{m=1}^{M} h_m(x_i)\]</div>
</div></blockquote>
<p>where the <span class="math notranslate nohighlight">\(h_m\)</span> are estimators called <em>weak learners</em> in the context
of boosting. Gradient Tree Boosting uses <a class="reference internal" href="tree.html#tree"><span class="std std-ref">decision tree regressors</span></a> of fixed size as weak learners. The constant M corresponds to the
<code class="docutils literal notranslate"><span class="pre">n_estimators</span></code> parameter.</p>
<p>Similar to other boosting algorithms, a GBRT is built in a greedy fashion:</p>
<blockquote>
<div><div class="math notranslate nohighlight">
\[F_m(x) = F_{m-1}(x) + h_m(x),\]</div>
</div></blockquote>
<p>where the newly added tree <span class="math notranslate nohighlight">\(h_m\)</span> is fitted in order to minimize a sum
of losses <span class="math notranslate nohighlight">\(L_m\)</span>, given the previous ensemble <span class="math notranslate nohighlight">\(F_{m-1}\)</span>:</p>
<blockquote>
<div><div class="math notranslate nohighlight">
\[h_m = \arg\min_{h} L_m = \arg\min_{h} \sum_{i=1}^{n}
l(y_i, F_{m-1}(x_i) + h(x_i)),\]</div>
</div></blockquote>
<p>where <span class="math notranslate nohighlight">\(l(y_i, F(x_i))\)</span> is defined by the <code class="docutils literal notranslate"><span class="pre">loss</span></code> parameter, detailed
in the next section.</p>
<p>By default, the initial model <span class="math notranslate nohighlight">\(F_{0}\)</span> is chosen as the constant that
minimizes the loss: for a least-squares loss, this is the empirical mean of
the target values. The initial model can also be specified via the <code class="docutils literal notranslate"><span class="pre">init</span></code>
argument.</p>
<p>Using a first-order Taylor approximation, the value of <span class="math notranslate nohighlight">\(l\)</span> can be
approximated as follows:</p>
<blockquote>
<div><div class="math notranslate nohighlight">
\[l(y_i, F_{m-1}(x_i) + h_m(x_i)) \approx
l(y_i, F_{m-1}(x_i))
+ h_m(x_i)
\left[ \frac{\partial l(y_i, F(x_i))}{\partial F(x_i)} \right]_{F=F_{m - 1}}.\]</div>
</div></blockquote>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>Briefly, a first-order Taylor approximation says that
<span class="math notranslate nohighlight">\(l(z) \approx l(a) + (z - a) \frac{\partial l}{\partial z}(a)\)</span>.
Here, <span class="math notranslate nohighlight">\(z\)</span> corresponds to <span class="math notranslate nohighlight">\(F_{m - 1}(x_i) + h_m(x_i)\)</span>, and
<span class="math notranslate nohighlight">\(a\)</span> corresponds to <span class="math notranslate nohighlight">\(F_{m-1}(x_i)\)</span></p>
</div>
<p>The quantity <span class="math notranslate nohighlight">\(\left[ \frac{\partial l(y_i, F(x_i))}{\partial F(x_i)}
\right]_{F=F_{m - 1}}\)</span> is the derivative of the loss with respect to its
second parameter, evaluated at <span class="math notranslate nohighlight">\(F_{m-1}(x)\)</span>. It is easy to compute for
any given <span class="math notranslate nohighlight">\(F_{m - 1}(x_i)\)</span> in a closed form since the loss is
differentiable. We will denote it by <span class="math notranslate nohighlight">\(g_i\)</span>.</p>
<p>Removing the constant terms, we have:</p>
<blockquote>
<div><div class="math notranslate nohighlight">
\[h_m \approx \arg\min_{h} \sum_{i=1}^{n} h(x_i) g_i\]</div>
</div></blockquote>
<p>This is minimized if <span class="math notranslate nohighlight">\(h(x_i)\)</span> is fitted to predict a value that is
proportional to the negative gradient <span class="math notranslate nohighlight">\(-g_i\)</span>. Therefore, at each
iteration, <strong>the estimator</strong> <span class="math notranslate nohighlight">\(h_m\)</span> <strong>is fitted to predict the negative
gradients of the samples</strong>. The gradients are updated at each iteration.
This can be considered as some kind of gradient descent in a functional
space.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>For some losses, e.g. <code class="docutils literal notranslate"><span class="pre">'absolute_error'</span></code> where the gradients
are <span class="math notranslate nohighlight">\(\pm 1\)</span>, the values predicted by a fitted <span class="math notranslate nohighlight">\(h_m\)</span> are not
accurate enough: the tree can only output integer values. As a result, the
leaves values of the tree <span class="math notranslate nohighlight">\(h_m\)</span> are modified once the tree is
fitted, such that the leaves values minimize the loss <span class="math notranslate nohighlight">\(L_m\)</span>. The
update is loss-dependent: for the absolute error loss, the value of
a leaf is updated to the median of the samples in that leaf.</p>
</div>
</section>
<section id="id11">
<h5><span class="section-number">1.11.1.2.5.2. </span>Classification<a class="headerlink" href="#id11" title="Permalink to this heading">¶</a></h5>
<p>Gradient boosting for classification is very similar to the regression case.
However, the sum of the trees <span class="math notranslate nohighlight">\(F_M(x_i) = \sum_m h_m(x_i)\)</span> is not
homogeneous to a prediction: it cannot be a class, since the trees predict
continuous values.</p>
<p>The mapping from the value <span class="math notranslate nohighlight">\(F_M(x_i)\)</span> to a class or a probability is
loss-dependent. For the log-loss, the probability that
<span class="math notranslate nohighlight">\(x_i\)</span> belongs to the positive class is modeled as <span class="math notranslate nohighlight">\(p(y_i = 1 |
x_i) = \sigma(F_M(x_i))\)</span> where <span class="math notranslate nohighlight">\(\sigma\)</span> is the sigmoid or expit function.</p>
<p>For multiclass classification, K trees (for K classes) are built at each of
the <span class="math notranslate nohighlight">\(M\)</span> iterations. The probability that <span class="math notranslate nohighlight">\(x_i\)</span> belongs to class
k is modeled as a softmax of the <span class="math notranslate nohighlight">\(F_{M,k}(x_i)\)</span> values.</p>
<p>Note that even for a classification task, the <span class="math notranslate nohighlight">\(h_m\)</span> sub-estimator is
still a regressor, not a classifier. This is because the sub-estimators are
trained to predict (negative) <em>gradients</em>, which are always continuous
quantities.</p>
</section>
</section>
<section id="loss-functions">
<span id="gradient-boosting-loss"></span><h4><span class="section-number">1.11.1.2.6. </span>Loss Functions<a class="headerlink" href="#loss-functions" title="Permalink to this heading">¶</a></h4>
<p>The following loss functions are supported and can be specified using
the parameter <code class="docutils literal notranslate"><span class="pre">loss</span></code>:</p>
<blockquote>
<div><ul class="simple">
<li><p>Regression</p>
<ul>
<li><p>Squared error (<code class="docutils literal notranslate"><span class="pre">'squared_error'</span></code>): The natural choice for regression
due to its superior computational properties. The initial model is
given by the mean of the target values.</p></li>
<li><p>Absolute error (<code class="docutils literal notranslate"><span class="pre">'absolute_error'</span></code>): A robust loss function for
regression. The initial model is given by the median of the
target values.</p></li>
<li><p>Huber (<code class="docutils literal notranslate"><span class="pre">'huber'</span></code>): Another robust loss function that combines
least squares and least absolute deviation; use <code class="docutils literal notranslate"><span class="pre">alpha</span></code> to
control the sensitivity with regards to outliers (see <a class="reference internal" href="#friedman2001" id="id12"><span>[Friedman2001]</span></a> for
more details).</p></li>
<li><p>Quantile (<code class="docutils literal notranslate"><span class="pre">'quantile'</span></code>): A loss function for quantile regression.
Use <code class="docutils literal notranslate"><span class="pre">0</span> <span class="pre"><</span> <span class="pre">alpha</span> <span class="pre"><</span> <span class="pre">1</span></code> to specify the quantile. This loss function
can be used to create prediction intervals
(see <a class="reference internal" href="../auto_examples/ensemble/plot_gradient_boosting_quantile.html#sphx-glr-auto-examples-ensemble-plot-gradient-boosting-quantile-py"><span class="std std-ref">Prediction Intervals for Gradient Boosting Regression</span></a>).</p></li>
</ul>
</li>
<li><p>Classification</p>
<ul>
<li><p>Binary log-loss (<code class="docutils literal notranslate"><span class="pre">'log-loss'</span></code>): The binomial
negative log-likelihood loss function for binary classification. It provides
probability estimates. The initial model is given by the
log odds-ratio.</p></li>
<li><p>Multi-class log-loss (<code class="docutils literal notranslate"><span class="pre">'log-loss'</span></code>): The multinomial
negative log-likelihood loss function for multi-class classification with
<code class="docutils literal notranslate"><span class="pre">n_classes</span></code> mutually exclusive classes. It provides
probability estimates. The initial model is given by the
prior probability of each class. At each iteration <code class="docutils literal notranslate"><span class="pre">n_classes</span></code>
regression trees have to be constructed which makes GBRT rather
inefficient for data sets with a large number of classes.</p></li>
<li><p>Exponential loss (<code class="docutils literal notranslate"><span class="pre">'exponential'</span></code>): The same loss function
as <a class="reference internal" href="generated/sklearn.ensemble.AdaBoostClassifier.html#sklearn.ensemble.AdaBoostClassifier" title="sklearn.ensemble.AdaBoostClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">AdaBoostClassifier</span></code></a>. Less robust to mislabeled
examples than <code class="docutils literal notranslate"><span class="pre">'log-loss'</span></code>; can only be used for binary
classification.</p></li>
</ul>
</li>
</ul>
</div></blockquote>
</section>
<section id="shrinkage-via-learning-rate">
<span id="gradient-boosting-shrinkage"></span><h4><span class="section-number">1.11.1.2.7. </span>Shrinkage via learning rate<a class="headerlink" href="#shrinkage-via-learning-rate" title="Permalink to this heading">¶</a></h4>
<p><a class="reference internal" href="#friedman2001" id="id13"><span>[Friedman2001]</span></a> proposed a simple regularization strategy that scales
the contribution of each weak learner by a constant factor <span class="math notranslate nohighlight">\(\nu\)</span>:</p>
<div class="math notranslate nohighlight">
\[F_m(x) = F_{m-1}(x) + \nu h_m(x)\]</div>
<p>The parameter <span class="math notranslate nohighlight">\(\nu\)</span> is also called the <strong>learning rate</strong> because
it scales the step length the gradient descent procedure; it can
be set via the <code class="docutils literal notranslate"><span class="pre">learning_rate</span></code> parameter.</p>
<p>The parameter <code class="docutils literal notranslate"><span class="pre">learning_rate</span></code> strongly interacts with the parameter
<code class="docutils literal notranslate"><span class="pre">n_estimators</span></code>, the number of weak learners to fit. Smaller values
of <code class="docutils literal notranslate"><span class="pre">learning_rate</span></code> require larger numbers of weak learners to maintain
a constant training error. Empirical evidence suggests that small
values of <code class="docutils literal notranslate"><span class="pre">learning_rate</span></code> favor better test error. <a class="reference internal" href="#htf" id="id14"><span>[HTF]</span></a>
recommend to set the learning rate to a small constant
(e.g. <code class="docutils literal notranslate"><span class="pre">learning_rate</span> <span class="pre"><=</span> <span class="pre">0.1</span></code>) and choose <code class="docutils literal notranslate"><span class="pre">n_estimators</span></code> by early
stopping. For a more detailed discussion of the interaction between
<code class="docutils literal notranslate"><span class="pre">learning_rate</span></code> and <code class="docutils literal notranslate"><span class="pre">n_estimators</span></code> see <a class="reference internal" href="#r2007" id="id15"><span>[R2007]</span></a>.</p>
</section>
<section id="subsampling">
<h4><span class="section-number">1.11.1.2.8. </span>Subsampling<a class="headerlink" href="#subsampling" title="Permalink to this heading">¶</a></h4>
<p><a class="reference internal" href="#friedman2002" id="id16"><span>[Friedman2002]</span></a> proposed stochastic gradient boosting, which combines gradient
boosting with bootstrap averaging (bagging). At each iteration
the base classifier is trained on a fraction <code class="docutils literal notranslate"><span class="pre">subsample</span></code> of
the available training data. The subsample is drawn without replacement.
A typical value of <code class="docutils literal notranslate"><span class="pre">subsample</span></code> is 0.5.</p>
<p>The figure below illustrates the effect of shrinkage and subsampling
on the goodness-of-fit of the model. We can clearly see that shrinkage
outperforms no-shrinkage. Subsampling with shrinkage can further increase
the accuracy of the model. Subsampling without shrinkage, on the other hand,
does poorly.</p>
<figure class="align-center">
<a class="reference external image-reference" href="../auto_examples/ensemble/plot_gradient_boosting_regularization.html"><img alt="../_images/sphx_glr_plot_gradient_boosting_regularization_001.png" src="../_images/sphx_glr_plot_gradient_boosting_regularization_001.png" style="width: 480.0px; height: 360.0px;" /></a>
</figure>
<p>Another strategy to reduce the variance is by subsampling the features
analogous to the random splits in <a class="reference internal" href="generated/sklearn.ensemble.RandomForestClassifier.html#sklearn.ensemble.RandomForestClassifier" title="sklearn.ensemble.RandomForestClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">RandomForestClassifier</span></code></a>.
The number of subsampled features can be controlled via the <code class="docutils literal notranslate"><span class="pre">max_features</span></code>
parameter.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>Using a small <code class="docutils literal notranslate"><span class="pre">max_features</span></code> value can significantly decrease the runtime.</p>
</div>
<p>Stochastic gradient boosting allows to compute out-of-bag estimates of the
test deviance by computing the improvement in deviance on the examples that are
not included in the bootstrap sample (i.e. the out-of-bag examples).
The improvements are stored in the attribute <code class="docutils literal notranslate"><span class="pre">oob_improvement_</span></code>.
<code class="docutils literal notranslate"><span class="pre">oob_improvement_[i]</span></code> holds the improvement in terms of the loss on the OOB samples
if you add the i-th stage to the current predictions.
Out-of-bag estimates can be used for model selection, for example to determine
the optimal number of iterations. OOB estimates are usually very pessimistic thus
we recommend to use cross-validation instead and only use OOB if cross-validation
is too time consuming.</p>
<aside class="topic">
<p class="topic-title">Examples:</p>
<ul class="simple">
<li><p><a class="reference internal" href="../auto_examples/ensemble/plot_gradient_boosting_regularization.html#sphx-glr-auto-examples-ensemble-plot-gradient-boosting-regularization-py"><span class="std std-ref">Gradient Boosting regularization</span></a></p></li>
<li><p><a class="reference internal" href="../auto_examples/ensemble/plot_gradient_boosting_oob.html#sphx-glr-auto-examples-ensemble-plot-gradient-boosting-oob-py"><span class="std std-ref">Gradient Boosting Out-of-Bag estimates</span></a></p></li>
<li><p><a class="reference internal" href="../auto_examples/ensemble/plot_ensemble_oob.html#sphx-glr-auto-examples-ensemble-plot-ensemble-oob-py"><span class="std std-ref">OOB Errors for Random Forests</span></a></p></li>
</ul>
</aside>
</section>
<section id="interpretation-with-feature-importance">
<h4><span class="section-number">1.11.1.2.9. </span>Interpretation with feature importance<a class="headerlink" href="#interpretation-with-feature-importance" title="Permalink to this heading">¶</a></h4>
<p>Individual decision trees can be interpreted easily by simply
visualizing the tree structure. Gradient boosting models, however,
comprise hundreds of regression trees thus they cannot be easily
interpreted by visual inspection of the individual trees. Fortunately,
a number of techniques have been proposed to summarize and interpret
gradient boosting models.</p>
<p>Often features do not contribute equally to predict the target
response; in many situations the majority of the features are in fact
irrelevant.
When interpreting a model, the first question usually is: what are
those important features and how do they contributing in predicting
the target response?</p>
<p>Individual decision trees intrinsically perform feature selection by selecting
appropriate split points. This information can be used to measure the
importance of each feature; the basic idea is: the more often a
feature is used in the split points of a tree the more important that
feature is. This notion of importance can be extended to decision tree
ensembles by simply averaging the impurity-based feature importance of each tree (see
<a class="reference internal" href="#random-forest-feature-importance"><span class="std std-ref">Feature importance evaluation</span></a> for more details).</p>
<p>The feature importance scores of a fit gradient boosting model can be
accessed via the <code class="docutils literal notranslate"><span class="pre">feature_importances_</span></code> property:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">sklearn.datasets</span> <span class="kn">import</span> <span class="n">make_hastie_10_2</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">sklearn.ensemble</span> <span class="kn">import</span> <span class="n">GradientBoostingClassifier</span>
<span class="gp">>>> </span><span class="n">X</span><span class="p">,</span> <span class="n">y</span> <span class="o">=</span> <span class="n">make_hastie_10_2</span><span class="p">(</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">clf</span> <span class="o">=</span> <span class="n">GradientBoostingClassifier</span><span class="p">(</span><span class="n">n_estimators</span><span class="o">=</span><span class="mi">100</span><span class="p">,</span> <span class="n">learning_rate</span><span class="o">=</span><span class="mf">1.0</span><span class="p">,</span>
<span class="gp">... </span> <span class="n">max_depth</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span> <span class="n">random_state</span><span class="o">=</span><span class="mi">0</span><span class="p">)</span><span class="o">.</span><span class="n">fit</span><span class="p">(</span><span class="n">X</span><span class="p">,</span> <span class="n">y</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">clf</span><span class="o">.</span><span class="n">feature_importances_</span>
<span class="go">array([0.10..., 0.10..., 0.11..., ...</span>
</pre></div>
</div>
<p>Note that this computation of feature importance is based on entropy, and it
is distinct from <a class="reference internal" href="generated/sklearn.inspection.permutation_importance.html#sklearn.inspection.permutation_importance" title="sklearn.inspection.permutation_importance"><code class="xref py py-func docutils literal notranslate"><span class="pre">sklearn.inspection.permutation_importance</span></code></a> which is
based on permutation of the features.</p>
<aside class="topic">
<p class="topic-title">Examples:</p>
<ul class="simple">