forked from scikit-learn/scikit-learn.github.io
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmanifold.html
834 lines (800 loc) · 52.3 KB
/
manifold.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
<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN"
"http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd">
<html xmlns="http://www.w3.org/1999/xhtml">
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8" />
<title>2.2. Manifold learning — scikit-learn 0.20.dev0 documentation</title>
<!-- htmltitle is before nature.css - we use this hack to load bootstrap first -->
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<link rel="stylesheet" href="../_static/css/bootstrap.min.css" media="screen" />
<link rel="stylesheet" href="../_static/css/bootstrap-responsive.css"/>
<link rel="stylesheet" href="../_static/nature.css" type="text/css" />
<link rel="stylesheet" href="../_static/pygments.css" type="text/css" />
<link rel="stylesheet" href="../_static/gallery.css" type="text/css" />
<script type="text/javascript">
var DOCUMENTATION_OPTIONS = {
URL_ROOT: '../',
VERSION: '0.20.dev0',
COLLAPSE_INDEX: false,
FILE_SUFFIX: '.html',
HAS_SOURCE: true,
SOURCELINK_SUFFIX: '.txt'
};
</script>
<script type="text/javascript" src="../_static/jquery.js"></script>
<script type="text/javascript" src="../_static/underscore.js"></script>
<script type="text/javascript" src="../_static/doctools.js"></script>
<script type="text/javascript" src="../_static/js/copybutton.js"></script>
<script type="text/javascript" src="../_static/js/extra.js"></script>
<script type="text/javascript" src="https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.0/MathJax.js?config=TeX-AMS_SVG"></script>
<link rel="shortcut icon" href="../_static/favicon.ico"/>
<link rel="author" title="About these documents" href="../about.html" />
<link rel="search" title="Search" href="../search.html" />
<link rel="next" title="2.3. Clustering" href="clustering.html" />
<link rel="prev" title="2.1. Gaussian mixture models" href="mixture.html" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<script src="../_static/js/bootstrap.min.js" type="text/javascript"></script>
<script>
VERSION_SUBDIR = (function(groups) {
return groups ? groups[1] : null;
})(location.href.match(/^https?:\/\/scikit-learn.org\/([^\/]+)/));
</script>
<link rel="canonical" href="http://scikit-learn.org/stable/modules/manifold.html" />
<script type="text/javascript">
$("div.buttonNext, div.buttonPrevious").hover(
function () {
$(this).css('background-color', '#FF9C34');
},
function () {
$(this).css('background-color', '#A7D6E2');
}
);
function showMenu() {
var topNav = document.getElementById("scikit-navbar");
if (topNav.className === "navbar") {
topNav.className += " responsive";
} else {
topNav.className = "navbar";
}
};
</script>
</head>
<body>
<div class="header-wrapper">
<div class="header">
<p class="logo"><a href="../index.html">
<img src="../_static/scikit-learn-logo-small.png" alt="Logo"/>
</a>
</p><div class="navbar" id="scikit-navbar">
<ul>
<li><a href="../index.html">Home</a></li>
<li><a href="../install.html">Installation</a></li>
<li class="btn-li"><div class="btn-group">
<a href="../documentation.html">Documentation</a>
<a class="btn dropdown-toggle" data-toggle="dropdown">
<span class="caret"></span>
</a>
<ul class="dropdown-menu">
<li class="link-title">Scikit-learn <script>document.write(DOCUMENTATION_OPTIONS.VERSION + (VERSION_SUBDIR ? " (" + VERSION_SUBDIR + ")" : ""));</script></li>
<li><a href="../tutorial/index.html">Tutorials</a></li>
<li><a href="../user_guide.html">User guide</a></li>
<li><a href="classes.html">API</a></li>
<li><a href="../glossary.html">Glossary</a></li>
<li><a href="../faq.html">FAQ</a></li>
<li><a href="../developers/contributing.html">Contributing</a></li>
<li class="divider"></li>
<script>if (VERSION_SUBDIR != "stable") document.write('<li><a href="http://scikit-learn.org/stable/documentation.html">Stable version</a></li>')</script>
<script>if (VERSION_SUBDIR != "dev") document.write('<li><a href="http://scikit-learn.org/dev/documentation.html">Development version</a></li>')</script>
<li><a href="http://scikit-learn.org/dev/versions.html">All available versions</a></li>
<li><a href="../_downloads/scikit-learn-docs.pdf">PDF documentation</a></li>
</ul>
</div>
</li>
<li><a href="../auto_examples/index.html">Examples</a></li>
</ul>
<a href="javascript:void(0);" onclick="showMenu()">
<div class="nav-icon">
<div class="hamburger-line"></div>
<div class="hamburger-line"></div>
<div class="hamburger-line"></div>
</div>
</a>
<div class="search_form">
<div class="gcse-search" id="cse" style="width: 100%;"></div>
</div>
</div> <!-- end navbar --></div>
</div>
<!-- GitHub "fork me" ribbon -->
<a href="https://github.com/scikit-learn/scikit-learn">
<img class="fork-me"
style="position: absolute; top: 0; right: 0; border: 0;"
src="../_static/img/forkme.png"
alt="Fork me on GitHub" />
</a>
<div class="content-wrapper">
<div class="sphinxsidebar">
<div class="sphinxsidebarwrapper">
<div class="rel">
<div class="rellink">
<a href="mixture.html"
accesskey="P">Previous
<br/>
<span class="smallrellink">
2.1. Gaussian...
</span>
<span class="hiddenrellink">
2.1. Gaussian mixture models
</span>
</a>
</div>
<div class="spacer">
</div>
<div class="rellink">
<a href="clustering.html"
accesskey="N">Next
<br/>
<span class="smallrellink">
2.3. Clustering
</span>
<span class="hiddenrellink">
2.3. Clustering
</span>
</a>
</div>
<!-- Ad a link to the 'up' page -->
<div class="spacer">
</div>
<div class="rellink">
<a href="../unsupervised_learning.html">
Up
<br/>
<span class="smallrellink">
2. Unsupervis...
</span>
<span class="hiddenrellink">
2. Unsupervised learning
</span>
</a>
</div>
</div>
<p class="doc-version"><b>scikit-learn v0.20.dev0</b><br/>
<a href="http://scikit-learn.org/stable/support.html#documentation-resources">Other versions</a></p>
<p class="citing">Please <b><a href="../about.html#citing-scikit-learn" style="font-size: 110%;">cite us </a></b>if you use the software.</p>
<ul>
<li><a class="reference internal" href="#">2.2. Manifold learning</a><ul>
<li><a class="reference internal" href="#introduction">2.2.1. Introduction</a></li>
<li><a class="reference internal" href="#isomap">2.2.2. Isomap</a><ul>
<li><a class="reference internal" href="#complexity">2.2.2.1. Complexity</a></li>
</ul>
</li>
<li><a class="reference internal" href="#locally-linear-embedding">2.2.3. Locally Linear Embedding</a><ul>
<li><a class="reference internal" href="#id3">2.2.3.1. Complexity</a></li>
</ul>
</li>
<li><a class="reference internal" href="#modified-locally-linear-embedding">2.2.4. Modified Locally Linear Embedding</a><ul>
<li><a class="reference internal" href="#id4">2.2.4.1. Complexity</a></li>
</ul>
</li>
<li><a class="reference internal" href="#hessian-eigenmapping">2.2.5. Hessian Eigenmapping</a><ul>
<li><a class="reference internal" href="#id5">2.2.5.1. Complexity</a></li>
</ul>
</li>
<li><a class="reference internal" href="#spectral-embedding">2.2.6. Spectral Embedding</a><ul>
<li><a class="reference internal" href="#id7">2.2.6.1. Complexity</a></li>
</ul>
</li>
<li><a class="reference internal" href="#local-tangent-space-alignment">2.2.7. Local Tangent Space Alignment</a><ul>
<li><a class="reference internal" href="#id8">2.2.7.1. Complexity</a></li>
</ul>
</li>
<li><a class="reference internal" href="#multi-dimensional-scaling-mds">2.2.8. Multi-dimensional Scaling (MDS)</a><ul>
<li><a class="reference internal" href="#metric-mds">2.2.8.1. Metric MDS</a></li>
<li><a class="reference internal" href="#nonmetric-mds">2.2.8.2. Nonmetric MDS</a></li>
</ul>
</li>
<li><a class="reference internal" href="#t-distributed-stochastic-neighbor-embedding-t-sne">2.2.9. t-distributed Stochastic Neighbor Embedding (t-SNE)</a><ul>
<li><a class="reference internal" href="#optimizing-t-sne">2.2.9.1. Optimizing t-SNE</a></li>
<li><a class="reference internal" href="#barnes-hut-t-sne">2.2.9.2. Barnes-Hut t-SNE</a></li>
</ul>
</li>
<li><a class="reference internal" href="#tips-on-practical-use">2.2.10. Tips on practical use</a></li>
</ul>
</li>
</ul>
</div>
</div>
<input type="checkbox" id="nav-trigger" class="nav-trigger" checked />
<label for="nav-trigger"></label>
<div class="content">
<div class="documentwrapper">
<div class="bodywrapper">
<div class="body" role="main">
<div class="section" id="manifold-learning">
<span id="manifold"></span><h1>2.2. Manifold learning<a class="headerlink" href="#manifold-learning" title="Permalink to this headline">¶</a></h1>
<div class="quote line-block">
<div class="line">Look for the bare necessities</div>
<div class="line">The simple bare necessities</div>
<div class="line">Forget about your worries and your strife</div>
<div class="line">I mean the bare necessities</div>
<div class="line">Old Mother Nature’s recipes</div>
<div class="line">That bring the bare necessities of life</div>
<div class="line"><br /></div>
<div class="line-block">
<div class="line">– Baloo’s song [The Jungle Book]</div>
</div>
</div>
<div class="figure align-center">
<a class="reference external image-reference" href="../auto_examples/manifold/plot_compare_methods.html"><img alt="../_images/sphx_glr_plot_compare_methods_0011.png" src="../_images/sphx_glr_plot_compare_methods_0011.png" style="width: 900.0px; height: 480.0px;" /></a>
</div>
<p>Manifold learning is an approach to non-linear dimensionality reduction.
Algorithms for this task are based on the idea that the dimensionality of
many data sets is only artificially high.</p>
<div class="section" id="introduction">
<h2>2.2.1. Introduction<a class="headerlink" href="#introduction" title="Permalink to this headline">¶</a></h2>
<p>High-dimensional datasets can be very difficult to visualize. While data
in two or three dimensions can be plotted to show the inherent
structure of the data, equivalent high-dimensional plots are much less
intuitive. To aid visualization of the structure of a dataset, the
dimension must be reduced in some way.</p>
<p>The simplest way to accomplish this dimensionality reduction is by taking
a random projection of the data. Though this allows some degree of
visualization of the data structure, the randomness of the choice leaves much
to be desired. In a random projection, it is likely that the more
interesting structure within the data will be lost.</p>
<p class="centered">
<strong><a class="reference external" href="../auto_examples/manifold/plot_lle_digits.html"><img alt="digits_img" src="../_images/sphx_glr_plot_lle_digits_0011.png" style="width: 320.0px; height: 240.0px;" /></a> <a class="reference external" href="../auto_examples/manifold/plot_lle_digits.html"><img alt="projected_img" src="../_images/sphx_glr_plot_lle_digits_0021.png" style="width: 320.0px; height: 240.0px;" /></a></strong></p><p>To address this concern, a number of supervised and unsupervised linear
dimensionality reduction frameworks have been designed, such as Principal
Component Analysis (PCA), Independent Component Analysis, Linear
Discriminant Analysis, and others. These algorithms define specific
rubrics to choose an “interesting” linear projection of the data.
These methods can be powerful, but often miss important non-linear
structure in the data.</p>
<p class="centered">
<strong><a class="reference external" href="../auto_examples/manifold/plot_lle_digits.html"><img alt="PCA_img" src="../_images/sphx_glr_plot_lle_digits_0031.png" style="width: 320.0px; height: 240.0px;" /></a> <a class="reference external" href="../auto_examples/manifold/plot_lle_digits.html"><img alt="LDA_img" src="../_images/sphx_glr_plot_lle_digits_0041.png" style="width: 320.0px; height: 240.0px;" /></a></strong></p><p>Manifold Learning can be thought of as an attempt to generalize linear
frameworks like PCA to be sensitive to non-linear structure in data. Though
supervised variants exist, the typical manifold learning problem is
unsupervised: it learns the high-dimensional structure of the data
from the data itself, without the use of predetermined classifications.</p>
<div class="topic">
<p class="topic-title first">Examples:</p>
<ul class="simple">
<li>See <a class="reference internal" href="../auto_examples/manifold/plot_lle_digits.html#sphx-glr-auto-examples-manifold-plot-lle-digits-py"><span class="std std-ref">Manifold learning on handwritten digits: Locally Linear Embedding, Isomap…</span></a> for an example of
dimensionality reduction on handwritten digits.</li>
<li>See <a class="reference internal" href="../auto_examples/manifold/plot_compare_methods.html#sphx-glr-auto-examples-manifold-plot-compare-methods-py"><span class="std std-ref">Comparison of Manifold Learning methods</span></a> for an example of
dimensionality reduction on a toy “S-curve” dataset.</li>
</ul>
</div>
<p>The manifold learning implementations available in scikit-learn are
summarized below</p>
</div>
<div class="section" id="isomap">
<span id="id1"></span><h2>2.2.2. Isomap<a class="headerlink" href="#isomap" title="Permalink to this headline">¶</a></h2>
<p>One of the earliest approaches to manifold learning is the Isomap
algorithm, short for Isometric Mapping. Isomap can be viewed as an
extension of Multi-dimensional Scaling (MDS) or Kernel PCA.
Isomap seeks a lower-dimensional embedding which maintains geodesic
distances between all points. Isomap can be performed with the object
<a class="reference internal" href="generated/sklearn.manifold.Isomap.html#sklearn.manifold.Isomap" title="sklearn.manifold.Isomap"><code class="xref py py-class docutils literal"><span class="pre">Isomap</span></code></a>.</p>
<div class="figure align-center">
<a class="reference external image-reference" href="../auto_examples/manifold/plot_lle_digits.html"><img alt="../_images/sphx_glr_plot_lle_digits_0051.png" src="../_images/sphx_glr_plot_lle_digits_0051.png" style="width: 320.0px; height: 240.0px;" /></a>
</div>
<div class="section" id="complexity">
<h3>2.2.2.1. Complexity<a class="headerlink" href="#complexity" title="Permalink to this headline">¶</a></h3>
<p>The Isomap algorithm comprises three stages:</p>
<ol class="arabic simple">
<li><strong>Nearest neighbor search.</strong> Isomap uses
<a class="reference internal" href="generated/sklearn.neighbors.BallTree.html#sklearn.neighbors.BallTree" title="sklearn.neighbors.BallTree"><code class="xref py py-class docutils literal"><span class="pre">sklearn.neighbors.BallTree</span></code></a> for efficient neighbor search.
The cost is approximately <span class="math">\(O[D \log(k) N \log(N)]\)</span>, for <span class="math">\(k\)</span>
nearest neighbors of <span class="math">\(N\)</span> points in <span class="math">\(D\)</span> dimensions.</li>
<li><strong>Shortest-path graph search.</strong> The most efficient known algorithms
for this are <em>Dijkstra’s Algorithm</em>, which is approximately
<span class="math">\(O[N^2(k + \log(N))]\)</span>, or the <em>Floyd-Warshall algorithm</em>, which
is <span class="math">\(O[N^3]\)</span>. The algorithm can be selected by the user with
the <code class="docutils literal"><span class="pre">path_method</span></code> keyword of <code class="docutils literal"><span class="pre">Isomap</span></code>. If unspecified, the code
attempts to choose the best algorithm for the input data.</li>
<li><strong>Partial eigenvalue decomposition.</strong> The embedding is encoded in the
eigenvectors corresponding to the <span class="math">\(d\)</span> largest eigenvalues of the
<span class="math">\(N \times N\)</span> isomap kernel. For a dense solver, the cost is
approximately <span class="math">\(O[d N^2]\)</span>. This cost can often be improved using
the <code class="docutils literal"><span class="pre">ARPACK</span></code> solver. The eigensolver can be specified by the user
with the <code class="docutils literal"><span class="pre">path_method</span></code> keyword of <code class="docutils literal"><span class="pre">Isomap</span></code>. If unspecified, the
code attempts to choose the best algorithm for the input data.</li>
</ol>
<p>The overall complexity of Isomap is
<span class="math">\(O[D \log(k) N \log(N)] + O[N^2(k + \log(N))] + O[d N^2]\)</span>.</p>
<ul class="simple">
<li><span class="math">\(N\)</span> : number of training data points</li>
<li><span class="math">\(D\)</span> : input dimension</li>
<li><span class="math">\(k\)</span> : number of nearest neighbors</li>
<li><span class="math">\(d\)</span> : output dimension</li>
</ul>
<div class="topic">
<p class="topic-title first">References:</p>
<ul class="simple">
<li><a class="reference external" href="http://science.sciencemag.org/content/290/5500/2319.full">“A global geometric framework for nonlinear dimensionality reduction”</a>
Tenenbaum, J.B.; De Silva, V.; & Langford, J.C. Science 290 (5500)</li>
</ul>
</div>
</div>
</div>
<div class="section" id="locally-linear-embedding">
<span id="id2"></span><h2>2.2.3. Locally Linear Embedding<a class="headerlink" href="#locally-linear-embedding" title="Permalink to this headline">¶</a></h2>
<p>Locally linear embedding (LLE) seeks a lower-dimensional projection of the data
which preserves distances within local neighborhoods. It can be thought
of as a series of local Principal Component Analyses which are globally
compared to find the best non-linear embedding.</p>
<p>Locally linear embedding can be performed with function
<a class="reference internal" href="generated/sklearn.manifold.locally_linear_embedding.html#sklearn.manifold.locally_linear_embedding" title="sklearn.manifold.locally_linear_embedding"><code class="xref py py-func docutils literal"><span class="pre">locally_linear_embedding</span></code></a> or its object-oriented counterpart
<a class="reference internal" href="generated/sklearn.manifold.LocallyLinearEmbedding.html#sklearn.manifold.LocallyLinearEmbedding" title="sklearn.manifold.LocallyLinearEmbedding"><code class="xref py py-class docutils literal"><span class="pre">LocallyLinearEmbedding</span></code></a>.</p>
<div class="figure align-center">
<a class="reference external image-reference" href="../auto_examples/manifold/plot_lle_digits.html"><img alt="../_images/sphx_glr_plot_lle_digits_0061.png" src="../_images/sphx_glr_plot_lle_digits_0061.png" style="width: 320.0px; height: 240.0px;" /></a>
</div>
<div class="section" id="id3">
<h3>2.2.3.1. Complexity<a class="headerlink" href="#id3" title="Permalink to this headline">¶</a></h3>
<p>The standard LLE algorithm comprises three stages:</p>
<ol class="arabic simple">
<li><strong>Nearest Neighbors Search</strong>. See discussion under Isomap above.</li>
<li><strong>Weight Matrix Construction</strong>. <span class="math">\(O[D N k^3]\)</span>.
The construction of the LLE weight matrix involves the solution of a
<span class="math">\(k \times k\)</span> linear equation for each of the <span class="math">\(N\)</span> local
neighborhoods</li>
<li><strong>Partial Eigenvalue Decomposition</strong>. See discussion under Isomap above.</li>
</ol>
<p>The overall complexity of standard LLE is
<span class="math">\(O[D \log(k) N \log(N)] + O[D N k^3] + O[d N^2]\)</span>.</p>
<ul class="simple">
<li><span class="math">\(N\)</span> : number of training data points</li>
<li><span class="math">\(D\)</span> : input dimension</li>
<li><span class="math">\(k\)</span> : number of nearest neighbors</li>
<li><span class="math">\(d\)</span> : output dimension</li>
</ul>
<div class="topic">
<p class="topic-title first">References:</p>
<ul class="simple">
<li><a class="reference external" href="http://www.sciencemag.org/content/290/5500/2323.full">“Nonlinear dimensionality reduction by locally linear embedding”</a>
Roweis, S. & Saul, L. Science 290:2323 (2000)</li>
</ul>
</div>
</div>
</div>
<div class="section" id="modified-locally-linear-embedding">
<h2>2.2.4. Modified Locally Linear Embedding<a class="headerlink" href="#modified-locally-linear-embedding" title="Permalink to this headline">¶</a></h2>
<p>One well-known issue with LLE is the regularization problem. When the number
of neighbors is greater than the number of input dimensions, the matrix
defining each local neighborhood is rank-deficient. To address this, standard
LLE applies an arbitrary regularization parameter <span class="math">\(r\)</span>, which is chosen
relative to the trace of the local weight matrix. Though it can be shown
formally that as <span class="math">\(r \to 0\)</span>, the solution converges to the desired
embedding, there is no guarantee that the optimal solution will be found
for <span class="math">\(r > 0\)</span>. This problem manifests itself in embeddings which distort
the underlying geometry of the manifold.</p>
<p>One method to address the regularization problem is to use multiple weight
vectors in each neighborhood. This is the essence of <em>modified locally
linear embedding</em> (MLLE). MLLE can be performed with function
<a class="reference internal" href="generated/sklearn.manifold.locally_linear_embedding.html#sklearn.manifold.locally_linear_embedding" title="sklearn.manifold.locally_linear_embedding"><code class="xref py py-func docutils literal"><span class="pre">locally_linear_embedding</span></code></a> or its object-oriented counterpart
<a class="reference internal" href="generated/sklearn.manifold.LocallyLinearEmbedding.html#sklearn.manifold.LocallyLinearEmbedding" title="sklearn.manifold.LocallyLinearEmbedding"><code class="xref py py-class docutils literal"><span class="pre">LocallyLinearEmbedding</span></code></a>, with the keyword <code class="docutils literal"><span class="pre">method</span> <span class="pre">=</span> <span class="pre">'modified'</span></code>.
It requires <code class="docutils literal"><span class="pre">n_neighbors</span> <span class="pre">></span> <span class="pre">n_components</span></code>.</p>
<div class="figure align-center">
<a class="reference external image-reference" href="../auto_examples/manifold/plot_lle_digits.html"><img alt="../_images/sphx_glr_plot_lle_digits_0071.png" src="../_images/sphx_glr_plot_lle_digits_0071.png" style="width: 320.0px; height: 240.0px;" /></a>
</div>
<div class="section" id="id4">
<h3>2.2.4.1. Complexity<a class="headerlink" href="#id4" title="Permalink to this headline">¶</a></h3>
<p>The MLLE algorithm comprises three stages:</p>
<ol class="arabic simple">
<li><strong>Nearest Neighbors Search</strong>. Same as standard LLE</li>
<li><strong>Weight Matrix Construction</strong>. Approximately
<span class="math">\(O[D N k^3] + O[N (k-D) k^2]\)</span>. The first term is exactly equivalent
to that of standard LLE. The second term has to do with constructing the
weight matrix from multiple weights. In practice, the added cost of
constructing the MLLE weight matrix is relatively small compared to the
cost of steps 1 and 3.</li>
<li><strong>Partial Eigenvalue Decomposition</strong>. Same as standard LLE</li>
</ol>
<p>The overall complexity of MLLE is
<span class="math">\(O[D \log(k) N \log(N)] + O[D N k^3] + O[N (k-D) k^2] + O[d N^2]\)</span>.</p>
<ul class="simple">
<li><span class="math">\(N\)</span> : number of training data points</li>
<li><span class="math">\(D\)</span> : input dimension</li>
<li><span class="math">\(k\)</span> : number of nearest neighbors</li>
<li><span class="math">\(d\)</span> : output dimension</li>
</ul>
<div class="topic">
<p class="topic-title first">References:</p>
<ul class="simple">
<li><a class="reference external" href="http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.70.382">“MLLE: Modified Locally Linear Embedding Using Multiple Weights”</a>
Zhang, Z. & Wang, J.</li>
</ul>
</div>
</div>
</div>
<div class="section" id="hessian-eigenmapping">
<h2>2.2.5. Hessian Eigenmapping<a class="headerlink" href="#hessian-eigenmapping" title="Permalink to this headline">¶</a></h2>
<p>Hessian Eigenmapping (also known as Hessian-based LLE: HLLE) is another method
of solving the regularization problem of LLE. It revolves around a
hessian-based quadratic form at each neighborhood which is used to recover
the locally linear structure. Though other implementations note its poor
scaling with data size, <code class="docutils literal"><span class="pre">sklearn</span></code> implements some algorithmic
improvements which make its cost comparable to that of other LLE variants
for small output dimension. HLLE can be performed with function
<a class="reference internal" href="generated/sklearn.manifold.locally_linear_embedding.html#sklearn.manifold.locally_linear_embedding" title="sklearn.manifold.locally_linear_embedding"><code class="xref py py-func docutils literal"><span class="pre">locally_linear_embedding</span></code></a> or its object-oriented counterpart
<a class="reference internal" href="generated/sklearn.manifold.LocallyLinearEmbedding.html#sklearn.manifold.LocallyLinearEmbedding" title="sklearn.manifold.LocallyLinearEmbedding"><code class="xref py py-class docutils literal"><span class="pre">LocallyLinearEmbedding</span></code></a>, with the keyword <code class="docutils literal"><span class="pre">method</span> <span class="pre">=</span> <span class="pre">'hessian'</span></code>.
It requires <code class="docutils literal"><span class="pre">n_neighbors</span> <span class="pre">></span> <span class="pre">n_components</span> <span class="pre">*</span> <span class="pre">(n_components</span> <span class="pre">+</span> <span class="pre">3)</span> <span class="pre">/</span> <span class="pre">2</span></code>.</p>
<div class="figure align-center">
<a class="reference external image-reference" href="../auto_examples/manifold/plot_lle_digits.html"><img alt="../_images/sphx_glr_plot_lle_digits_0081.png" src="../_images/sphx_glr_plot_lle_digits_0081.png" style="width: 320.0px; height: 240.0px;" /></a>
</div>
<div class="section" id="id5">
<h3>2.2.5.1. Complexity<a class="headerlink" href="#id5" title="Permalink to this headline">¶</a></h3>
<p>The HLLE algorithm comprises three stages:</p>
<ol class="arabic simple">
<li><strong>Nearest Neighbors Search</strong>. Same as standard LLE</li>
<li><strong>Weight Matrix Construction</strong>. Approximately
<span class="math">\(O[D N k^3] + O[N d^6]\)</span>. The first term reflects a similar
cost to that of standard LLE. The second term comes from a QR
decomposition of the local hessian estimator.</li>
<li><strong>Partial Eigenvalue Decomposition</strong>. Same as standard LLE</li>
</ol>
<p>The overall complexity of standard HLLE is
<span class="math">\(O[D \log(k) N \log(N)] + O[D N k^3] + O[N d^6] + O[d N^2]\)</span>.</p>
<ul class="simple">
<li><span class="math">\(N\)</span> : number of training data points</li>
<li><span class="math">\(D\)</span> : input dimension</li>
<li><span class="math">\(k\)</span> : number of nearest neighbors</li>
<li><span class="math">\(d\)</span> : output dimension</li>
</ul>
<div class="topic">
<p class="topic-title first">References:</p>
<ul class="simple">
<li><a class="reference external" href="http://www.pnas.org/content/100/10/5591">“Hessian Eigenmaps: Locally linear embedding techniques for
high-dimensional data”</a>
Donoho, D. & Grimes, C. Proc Natl Acad Sci USA. 100:5591 (2003)</li>
</ul>
</div>
</div>
</div>
<div class="section" id="spectral-embedding">
<span id="id6"></span><h2>2.2.6. Spectral Embedding<a class="headerlink" href="#spectral-embedding" title="Permalink to this headline">¶</a></h2>
<p>Spectral Embedding is an approach to calculating a non-linear embedding.
Scikit-learn implements Laplacian Eigenmaps, which finds a low dimensional
representation of the data using a spectral decomposition of the graph
Laplacian. The graph generated can be considered as a discrete approximation of
the low dimensional manifold in the high dimensional space. Minimization of a
cost function based on the graph ensures that points close to each other on
the manifold are mapped close to each other in the low dimensional space,
preserving local distances. Spectral embedding can be performed with the
function <a class="reference internal" href="generated/sklearn.manifold.spectral_embedding.html#sklearn.manifold.spectral_embedding" title="sklearn.manifold.spectral_embedding"><code class="xref py py-func docutils literal"><span class="pre">spectral_embedding</span></code></a> or its object-oriented counterpart
<a class="reference internal" href="generated/sklearn.manifold.SpectralEmbedding.html#sklearn.manifold.SpectralEmbedding" title="sklearn.manifold.SpectralEmbedding"><code class="xref py py-class docutils literal"><span class="pre">SpectralEmbedding</span></code></a>.</p>
<div class="section" id="id7">
<h3>2.2.6.1. Complexity<a class="headerlink" href="#id7" title="Permalink to this headline">¶</a></h3>
<p>The Spectral Embedding (Laplacian Eigenmaps) algorithm comprises three stages:</p>
<ol class="arabic simple">
<li><strong>Weighted Graph Construction</strong>. Transform the raw input data into
graph representation using affinity (adjacency) matrix representation.</li>
<li><strong>Graph Laplacian Construction</strong>. unnormalized Graph Laplacian
is constructed as <span class="math">\(L = D - A\)</span> for and normalized one as
<span class="math">\(L = D^{-\frac{1}{2}} (D - A) D^{-\frac{1}{2}}\)</span>.</li>
<li><strong>Partial Eigenvalue Decomposition</strong>. Eigenvalue decomposition is
done on graph Laplacian</li>
</ol>
<p>The overall complexity of spectral embedding is
<span class="math">\(O[D \log(k) N \log(N)] + O[D N k^3] + O[d N^2]\)</span>.</p>
<ul class="simple">
<li><span class="math">\(N\)</span> : number of training data points</li>
<li><span class="math">\(D\)</span> : input dimension</li>
<li><span class="math">\(k\)</span> : number of nearest neighbors</li>
<li><span class="math">\(d\)</span> : output dimension</li>
</ul>
<div class="topic">
<p class="topic-title first">References:</p>
<ul class="simple">
<li><a class="reference external" href="http://web.cse.ohio-state.edu/~mbelkin/papers/LEM_NC_03.pdf">“Laplacian Eigenmaps for Dimensionality Reduction
and Data Representation”</a>
M. Belkin, P. Niyogi, Neural Computation, June 2003; 15 (6):1373-1396</li>
</ul>
</div>
</div>
</div>
<div class="section" id="local-tangent-space-alignment">
<h2>2.2.7. Local Tangent Space Alignment<a class="headerlink" href="#local-tangent-space-alignment" title="Permalink to this headline">¶</a></h2>
<p>Though not technically a variant of LLE, Local tangent space alignment (LTSA)
is algorithmically similar enough to LLE that it can be put in this category.
Rather than focusing on preserving neighborhood distances as in LLE, LTSA
seeks to characterize the local geometry at each neighborhood via its
tangent space, and performs a global optimization to align these local
tangent spaces to learn the embedding. LTSA can be performed with function
<a class="reference internal" href="generated/sklearn.manifold.locally_linear_embedding.html#sklearn.manifold.locally_linear_embedding" title="sklearn.manifold.locally_linear_embedding"><code class="xref py py-func docutils literal"><span class="pre">locally_linear_embedding</span></code></a> or its object-oriented counterpart
<a class="reference internal" href="generated/sklearn.manifold.LocallyLinearEmbedding.html#sklearn.manifold.LocallyLinearEmbedding" title="sklearn.manifold.LocallyLinearEmbedding"><code class="xref py py-class docutils literal"><span class="pre">LocallyLinearEmbedding</span></code></a>, with the keyword <code class="docutils literal"><span class="pre">method</span> <span class="pre">=</span> <span class="pre">'ltsa'</span></code>.</p>
<div class="figure align-center">
<a class="reference external image-reference" href="../auto_examples/manifold/plot_lle_digits.html"><img alt="../_images/sphx_glr_plot_lle_digits_0091.png" src="../_images/sphx_glr_plot_lle_digits_0091.png" style="width: 320.0px; height: 240.0px;" /></a>
</div>
<div class="section" id="id8">
<h3>2.2.7.1. Complexity<a class="headerlink" href="#id8" title="Permalink to this headline">¶</a></h3>
<p>The LTSA algorithm comprises three stages:</p>
<ol class="arabic simple">
<li><strong>Nearest Neighbors Search</strong>. Same as standard LLE</li>
<li><strong>Weight Matrix Construction</strong>. Approximately
<span class="math">\(O[D N k^3] + O[k^2 d]\)</span>. The first term reflects a similar
cost to that of standard LLE.</li>
<li><strong>Partial Eigenvalue Decomposition</strong>. Same as standard LLE</li>
</ol>
<p>The overall complexity of standard LTSA is
<span class="math">\(O[D \log(k) N \log(N)] + O[D N k^3] + O[k^2 d] + O[d N^2]\)</span>.</p>
<ul class="simple">
<li><span class="math">\(N\)</span> : number of training data points</li>
<li><span class="math">\(D\)</span> : input dimension</li>
<li><span class="math">\(k\)</span> : number of nearest neighbors</li>
<li><span class="math">\(d\)</span> : output dimension</li>
</ul>
<div class="topic">
<p class="topic-title first">References:</p>
<ul class="simple">
<li><a class="reference external" href="http://citeseer.ist.psu.edu/viewdoc/summary?doi=10.1.1.4.3693">“Principal manifolds and nonlinear dimensionality reduction via
tangent space alignment”</a>
Zhang, Z. & Zha, H. Journal of Shanghai Univ. 8:406 (2004)</li>
</ul>
</div>
</div>
</div>
<div class="section" id="multi-dimensional-scaling-mds">
<span id="multidimensional-scaling"></span><h2>2.2.8. Multi-dimensional Scaling (MDS)<a class="headerlink" href="#multi-dimensional-scaling-mds" title="Permalink to this headline">¶</a></h2>
<p><a class="reference external" href="https://en.wikipedia.org/wiki/Multidimensional_scaling">Multidimensional scaling</a>
(<a class="reference internal" href="generated/sklearn.manifold.MDS.html#sklearn.manifold.MDS" title="sklearn.manifold.MDS"><code class="xref py py-class docutils literal"><span class="pre">MDS</span></code></a>) seeks a low-dimensional
representation of the data in which the distances respect well the
distances in the original high-dimensional space.</p>
<p>In general, is a technique used for analyzing similarity or
dissimilarity data. <a class="reference internal" href="generated/sklearn.manifold.MDS.html#sklearn.manifold.MDS" title="sklearn.manifold.MDS"><code class="xref py py-class docutils literal"><span class="pre">MDS</span></code></a> attempts to model similarity or dissimilarity data as
distances in a geometric spaces. The data can be ratings of similarity between
objects, interaction frequencies of molecules, or trade indices between
countries.</p>
<p>There exists two types of MDS algorithm: metric and non metric. In the
scikit-learn, the class <a class="reference internal" href="generated/sklearn.manifold.MDS.html#sklearn.manifold.MDS" title="sklearn.manifold.MDS"><code class="xref py py-class docutils literal"><span class="pre">MDS</span></code></a> implements both. In Metric MDS, the input
similarity matrix arises from a metric (and thus respects the triangular
inequality), the distances between output two points are then set to be as
close as possible to the similarity or dissimilarity data. In the non-metric
version, the algorithms will try to preserve the order of the distances, and
hence seek for a monotonic relationship between the distances in the embedded
space and the similarities/dissimilarities.</p>
<div class="figure align-center">
<a class="reference external image-reference" href="../auto_examples/manifold/plot_lle_digits.html"><img alt="../_images/sphx_glr_plot_lle_digits_0101.png" src="../_images/sphx_glr_plot_lle_digits_0101.png" style="width: 320.0px; height: 240.0px;" /></a>
</div>
<p>Let <span class="math">\(S\)</span> be the similarity matrix, and <span class="math">\(X\)</span> the coordinates of the
<span class="math">\(n\)</span> input points. Disparities <span class="math">\(\hat{d}_{ij}\)</span> are transformation of
the similarities chosen in some optimal ways. The objective, called the
stress, is then defined by <span class="math">\(sum_{i < j} d_{ij}(X) - \hat{d}_{ij}(X)\)</span></p>
<div class="section" id="metric-mds">
<h3>2.2.8.1. Metric MDS<a class="headerlink" href="#metric-mds" title="Permalink to this headline">¶</a></h3>
<p>The simplest metric <a class="reference internal" href="generated/sklearn.manifold.MDS.html#sklearn.manifold.MDS" title="sklearn.manifold.MDS"><code class="xref py py-class docutils literal"><span class="pre">MDS</span></code></a> model, called <em>absolute MDS</em>, disparities are defined by
<span class="math">\(\hat{d}_{ij} = S_{ij}\)</span>. With absolute MDS, the value <span class="math">\(S_{ij}\)</span>
should then correspond exactly to the distance between point <span class="math">\(i\)</span> and
<span class="math">\(j\)</span> in the embedding point.</p>
<p>Most commonly, disparities are set to <span class="math">\(\hat{d}_{ij} = b S_{ij}\)</span>.</p>
</div>
<div class="section" id="nonmetric-mds">
<h3>2.2.8.2. Nonmetric MDS<a class="headerlink" href="#nonmetric-mds" title="Permalink to this headline">¶</a></h3>
<p>Non metric <a class="reference internal" href="generated/sklearn.manifold.MDS.html#sklearn.manifold.MDS" title="sklearn.manifold.MDS"><code class="xref py py-class docutils literal"><span class="pre">MDS</span></code></a> focuses on the ordination of the data. If
<span class="math">\(S_{ij} < S_{kl}\)</span>, then the embedding should enforce <span class="math">\(d_{ij} <
d_{jk}\)</span>. A simple algorithm to enforce that is to use a monotonic regression
of <span class="math">\(d_{ij}\)</span> on <span class="math">\(S_{ij}\)</span>, yielding disparities <span class="math">\(\hat{d}_{ij}\)</span>
in the same order as <span class="math">\(S_{ij}\)</span>.</p>
<p>A trivial solution to this problem is to set all the points on the origin. In
order to avoid that, the disparities <span class="math">\(\hat{d}_{ij}\)</span> are normalized.</p>
<div class="figure align-center">
<a class="reference external image-reference" href="../auto_examples/manifold/plot_mds.html"><img alt="../_images/sphx_glr_plot_mds_0011.png" src="../_images/sphx_glr_plot_mds_0011.png" style="width: 384.0px; height: 288.0px;" /></a>
</div>
<div class="topic">
<p class="topic-title first">References:</p>
<ul class="simple">
<li><a class="reference external" href="http://www.springer.com/fr/book/9780387251509">“Modern Multidimensional Scaling - Theory and Applications”</a>
Borg, I.; Groenen P. Springer Series in Statistics (1997)</li>
<li><a class="reference external" href="http://link.springer.com/article/10.1007%2FBF02289694">“Nonmetric multidimensional scaling: a numerical method”</a>
Kruskal, J. Psychometrika, 29 (1964)</li>
<li><a class="reference external" href="http://link.springer.com/article/10.1007%2FBF02289565">“Multidimensional scaling by optimizing goodness of fit to a nonmetric hypothesis”</a>
Kruskal, J. Psychometrika, 29, (1964)</li>
</ul>
</div>
</div>
</div>
<div class="section" id="t-distributed-stochastic-neighbor-embedding-t-sne">
<span id="t-sne"></span><h2>2.2.9. t-distributed Stochastic Neighbor Embedding (t-SNE)<a class="headerlink" href="#t-distributed-stochastic-neighbor-embedding-t-sne" title="Permalink to this headline">¶</a></h2>
<p>t-SNE (<a class="reference internal" href="generated/sklearn.manifold.TSNE.html#sklearn.manifold.TSNE" title="sklearn.manifold.TSNE"><code class="xref py py-class docutils literal"><span class="pre">TSNE</span></code></a>) converts affinities of data points to probabilities.
The affinities in the original space are represented by Gaussian joint
probabilities and the affinities in the embedded space are represented by
Student’s t-distributions. This allows t-SNE to be particularly sensitive
to local structure and has a few other advantages over existing techniques:</p>
<ul class="simple">
<li>Revealing the structure at many scales on a single map</li>
<li>Revealing data that lie in multiple, different, manifolds or clusters</li>
<li>Reducing the tendency to crowd points together at the center</li>
</ul>
<p>While Isomap, LLE and variants are best suited to unfold a single continuous
low dimensional manifold, t-SNE will focus on the local structure of the data
and will tend to extract clustered local groups of samples as highlighted on
the S-curve example. This ability to group samples based on the local structure
might be beneficial to visually disentangle a dataset that comprises several
manifolds at once as is the case in the digits dataset.</p>
<p>The Kullback-Leibler (KL) divergence of the joint
probabilities in the original space and the embedded space will be minimized
by gradient descent. Note that the KL divergence is not convex, i.e.
multiple restarts with different initializations will end up in local minima
of the KL divergence. Hence, it is sometimes useful to try different seeds
and select the embedding with the lowest KL divergence.</p>
<p>The disadvantages to using t-SNE are roughly:</p>
<ul class="simple">
<li>t-SNE is computationally expensive, and can take several hours on million-sample
datasets where PCA will finish in seconds or minutes</li>
<li>The Barnes-Hut t-SNE method is limited to two or three dimensional embeddings.</li>
<li>The algorithm is stochastic and multiple restarts with different seeds can
yield different embeddings. However, it is perfectly legitimate to pick the
embedding with the least error.</li>
<li>Global structure is not explicitly preserved. This is problem is mitigated by
initializing points with PCA (using <cite>init=’pca’</cite>).</li>
</ul>
<div class="figure align-center">
<a class="reference external image-reference" href="../auto_examples/manifold/plot_lle_digits.html"><img alt="../_images/sphx_glr_plot_lle_digits_0131.png" src="../_images/sphx_glr_plot_lle_digits_0131.png" style="width: 320.0px; height: 240.0px;" /></a>
</div>
<div class="section" id="optimizing-t-sne">
<h3>2.2.9.1. Optimizing t-SNE<a class="headerlink" href="#optimizing-t-sne" title="Permalink to this headline">¶</a></h3>
<p>The main purpose of t-SNE is visualization of high-dimensional data. Hence,
it works best when the data will be embedded on two or three dimensions.</p>
<p>Optimizing the KL divergence can be a little bit tricky sometimes. There are
five parameters that control the optimization of t-SNE and therefore possibly
the quality of the resulting embedding:</p>
<ul class="simple">
<li>perplexity</li>
<li>early exaggeration factor</li>
<li>learning rate</li>
<li>maximum number of iterations</li>
<li>angle (not used in the exact method)</li>
</ul>
<p>The perplexity is defined as <span class="math">\(k=2^{(S)}\)</span> where <span class="math">\(S\)</span> is the Shannon
entropy of the conditional probability distribution. The perplexity of a
<span class="math">\(k\)</span>-sided die is <span class="math">\(k\)</span>, so that <span class="math">\(k\)</span> is effectively the number of
nearest neighbors t-SNE considers when generating the conditional probabilities.
Larger perplexities lead to more nearest neighbors and less sensitive to small
structure. Conversely a lower perplexity considers a smaller number of
neighbors, and thus ignores more global information in favour of the
local neighborhood. As dataset sizes get larger more points will be
required to get a reasonable sample of the local neighborhood, and hence
larger perplexities may be required. Similarly noisier datasets will require
larger perplexity values to encompass enough local neighbors to see beyond
the background noise.</p>
<p>The maximum number of iterations is usually high enough and does not need
any tuning. The optimization consists of two phases: the early exaggeration
phase and the final optimization. During early exaggeration the joint
probabilities in the original space will be artificially increased by
multiplication with a given factor. Larger factors result in larger gaps
between natural clusters in the data. If the factor is too high, the KL
divergence could increase during this phase. Usually it does not have to be
tuned. A critical parameter is the learning rate. If it is too low gradient
descent will get stuck in a bad local minimum. If it is too high the KL
divergence will increase during optimization. More tips can be found in
Laurens van der Maaten’s FAQ (see references). The last parameter, angle,
is a tradeoff between performance and accuracy. Larger angles imply that we
can approximate larger regions by a single point, leading to better speed
but less accurate results.</p>
<p><a class="reference external" href="http://distill.pub/2016/misread-tsne/">“How to Use t-SNE Effectively”</a>
provides a good discussion of the effects of the various parameters, as well
as interactive plots to explore the effects of different parameters.</p>
</div>
<div class="section" id="barnes-hut-t-sne">
<h3>2.2.9.2. Barnes-Hut t-SNE<a class="headerlink" href="#barnes-hut-t-sne" title="Permalink to this headline">¶</a></h3>
<p>The Barnes-Hut t-SNE that has been implemented here is usually much slower than
other manifold learning algorithms. The optimization is quite difficult
and the computation of the gradient is <span class="math">\(O[d N log(N)]\)</span>, where <span class="math">\(d\)</span>
is the number of output dimensions and <span class="math">\(N\)</span> is the number of samples. The
Barnes-Hut method improves on the exact method where t-SNE complexity is
<span class="math">\(O[d N^2]\)</span>, but has several other notable differences:</p>
<ul class="simple">
<li>The Barnes-Hut implementation only works when the target dimensionality is 3
or less. The 2D case is typical when building visualizations.</li>
<li>Barnes-Hut only works with dense input data. Sparse data matrices can only be
embedded with the exact method or can be approximated by a dense low rank
projection for instance using <a class="reference internal" href="generated/sklearn.decomposition.TruncatedSVD.html#sklearn.decomposition.TruncatedSVD" title="sklearn.decomposition.TruncatedSVD"><code class="xref py py-class docutils literal"><span class="pre">sklearn.decomposition.TruncatedSVD</span></code></a></li>
<li>Barnes-Hut is an approximation of the exact method. The approximation is
parameterized with the angle parameter, therefore the angle parameter is
unused when method=”exact”</li>
<li>Barnes-Hut is significantly more scalable. Barnes-Hut can be used to embed
hundred of thousands of data points while the exact method can handle
thousands of samples before becoming computationally intractable</li>
</ul>
<p>For visualization purpose (which is the main use case of t-SNE), using the
Barnes-Hut method is strongly recommended. The exact t-SNE method is useful
for checking the theoretically properties of the embedding possibly in higher
dimensional space but limit to small datasets due to computational constraints.</p>
<p>Also note that the digits labels roughly match the natural grouping found by
t-SNE while the linear 2D projection of the PCA model yields a representation
where label regions largely overlap. This is a strong clue that this data can
be well separated by non linear methods that focus on the local structure (e.g.
an SVM with a Gaussian RBF kernel). However, failing to visualize well
separated homogeneously labeled groups with t-SNE in 2D does not necessarily
imply that the data cannot be correctly classified by a supervised model. It
might be the case that 2 dimensions are not low enough to accurately represents
the internal structure of the data.</p>
<div class="topic">
<p class="topic-title first">References:</p>
<ul class="simple">
<li><a class="reference external" href="http://jmlr.org/papers/v9/vandermaaten08a.html">“Visualizing High-Dimensional Data Using t-SNE”</a>
van der Maaten, L.J.P.; Hinton, G. Journal of Machine Learning Research
(2008)</li>
<li><a class="reference external" href="http://lvdmaaten.github.io/tsne/">“t-Distributed Stochastic Neighbor Embedding”</a>
van der Maaten, L.J.P.</li>
<li><a class="reference external" href="https://lvdmaaten.github.io/publications/papers/JMLR_2014.pdf">“Accelerating t-SNE using Tree-Based Algorithms.”</a>
L.J.P. van der Maaten. Journal of Machine Learning Research 15(Oct):3221-3245, 2014.</li>
</ul>
</div>
</div>
</div>
<div class="section" id="tips-on-practical-use">
<h2>2.2.10. Tips on practical use<a class="headerlink" href="#tips-on-practical-use" title="Permalink to this headline">¶</a></h2>
<ul class="simple">
<li>Make sure the same scale is used over all features. Because manifold
learning methods are based on a nearest-neighbor search, the algorithm
may perform poorly otherwise. See <a class="reference internal" href="preprocessing.html#preprocessing-scaler"><span class="std std-ref">StandardScaler</span></a>
for convenient ways of scaling heterogeneous data.</li>
<li>The reconstruction error computed by each routine can be used to choose
the optimal output dimension. For a <span class="math">\(d\)</span>-dimensional manifold embedded
in a <span class="math">\(D\)</span>-dimensional parameter space, the reconstruction error will
decrease as <code class="docutils literal"><span class="pre">n_components</span></code> is increased until <code class="docutils literal"><span class="pre">n_components</span> <span class="pre">==</span> <span class="pre">d</span></code>.</li>
<li>Note that noisy data can “short-circuit” the manifold, in essence acting
as a bridge between parts of the manifold that would otherwise be
well-separated. Manifold learning on noisy and/or incomplete data is
an active area of research.</li>
<li>Certain input configurations can lead to singular weight matrices, for
example when more than two points in the dataset are identical, or when
the data is split into disjointed groups. In this case, <code class="docutils literal"><span class="pre">solver='arpack'</span></code>
will fail to find the null space. The easiest way to address this is to
use <code class="docutils literal"><span class="pre">solver='dense'</span></code> which will work on a singular matrix, though it may
be very slow depending on the number of input points. Alternatively, one
can attempt to understand the source of the singularity: if it is due to
disjoint sets, increasing <code class="docutils literal"><span class="pre">n_neighbors</span></code> may help. If it is due to
identical points in the dataset, removing these points may help.</li>
</ul>
<div class="admonition seealso">
<p class="first admonition-title">See also</p>
<p class="last"><a class="reference internal" href="ensemble.html#random-trees-embedding"><span class="std std-ref">Totally Random Trees Embedding</span></a> can also be useful to derive non-linear
representations of feature space, also it does not perform
dimensionality reduction.</p>
</div>
</div>
</div>
</div>
</div>
</div>
<div class="clearer"></div>
</div>
</div>
<div class="footer">
© 2007 - 2018, scikit-learn developers (BSD License).
<a href="../_sources/modules/manifold.rst.txt" rel="nofollow">Show this page source</a>
</div>
<div class="rel">
<div class="buttonPrevious">
<a href="mixture.html">Previous
</a>
</div>
<div class="buttonNext">
<a href="clustering.html">Next
</a>
</div>
</div>
<script type="text/javascript">
var _gaq = _gaq || [];
_gaq.push(['_setAccount', 'UA-22606712-2']);
_gaq.push(['_trackPageview']);
(function() {
var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true;
ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js';
var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s);
})();
</script>
<script>
(function() {
var cx = '016639176250731907682:tjtqbvtvij0';
var gcse = document.createElement('script'); gcse.type = 'text/javascript'; gcse.async = true;
gcse.src = 'https://cse.google.com/cse.js?cx=' + cx;
var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(gcse, s);
})();
</script>
</body>
</html>