1
00:00:00,090 --> 00:00:02,320
In this video, I'd like to start to talk about clustering.

2
00:00:03,420 --> 00:00:04,850
This will be exciting because this

3
00:00:04,930 --> 00:00:06,910
is our first unsupervised learning algorithm

4
00:00:07,350 --> 00:00:09,080
where we learn from unlabeled data

5
00:00:09,840 --> 00:00:10,730
instead of the label data.

6
00:00:11,900 --> 00:00:13,300
So, what is unsupervised learning?

7
00:00:14,390 --> 00:00:15,630
I briefly talked about unsupervised

8
00:00:16,350 --> 00:00:17,470
learning at the beginning of

9
00:00:17,550 --> 00:00:18,560
the class but it's useful

10
00:00:19,030 --> 00:00:20,320
to contrast it with supervised learning. So, here's

11
00:00:21,760 --> 00:00:23,750
a typical supervisory problem where

12
00:00:24,030 --> 00:00:25,470
we are given a label training sets

13
00:00:25,770 --> 00:00:27,470
and the goal is to find

14
00:00:27,980 --> 00:00:29,420
the decision boundary that separates the

15
00:00:29,530 --> 00:00:31,310
positive label examples and the negative label examples.

16
00:00:33,100 --> 00:00:34,400
The supervised learning problem in

17
00:00:34,460 --> 00:00:35,710
this case is given a

18
00:00:35,850 --> 00:00:38,270
set of labels to fit a hypothesis to it.

19
00:00:39,160 --> 00:00:40,560
In contrast, in the unsupervised

20
00:00:41,080 --> 00:00:42,420
learning problem, we're given

21
00:00:42,710 --> 00:00:43,740
data that does not have

22
00:00:43,890 --> 00:00:45,270
any labels associated with it.

23
00:00:46,730 --> 00:00:47,940
So we're given data that looks

24
00:00:48,100 --> 00:00:49,090
like this, here's a set

25
00:00:49,180 --> 00:00:50,470
of points and then no labels.

26
00:00:51,800 --> 00:00:52,860
And so our training set is written

27
00:00:53,220 --> 00:00:54,720
just x1, x2 and

28
00:00:55,210 --> 00:00:56,890
so on up to x(m)

29
00:00:57,450 --> 00:00:58,720
and we don't get any labels y.

30
00:00:59,540 --> 00:01:00,800
And that's why the points plotted

31
00:01:01,160 --> 00:01:02,300
up on the figure don't have

32
00:01:02,430 --> 00:01:04,330
any labels on them.

33
00:01:04,490 --> 00:01:05,510
So in unsupervised learning, what

34
00:01:05,710 --> 00:01:06,860
we do is, we give this sort of

35
00:01:07,280 --> 00:01:09,150
unlabeled training set to

36
00:01:09,250 --> 00:01:10,510
an algorithm and we just

37
00:01:10,600 --> 00:01:12,220
ask the algorithm: find some

38
00:01:12,430 --> 00:01:14,130
structure in the data for us.

39
00:01:15,420 --> 00:01:16,490
Given this data set, one

40
00:01:16,650 --> 00:01:17,810
type of structure we might

41
00:01:18,010 --> 00:01:19,540
have an algorithm find, is that

42
00:01:19,810 --> 00:01:21,440
it looks like this data set has

43
00:01:21,620 --> 00:01:23,740
points grouped into two

44
00:01:24,030 --> 00:01:25,500
separate clusters and so

45
00:01:25,800 --> 00:01:28,230
an algorithm that finds that

46
00:01:28,360 --> 00:01:29,230
clusters like the ones I just

47
00:01:29,450 --> 00:01:30,610
circled, is called a clustering

48
00:01:32,440 --> 00:01:32,440
algorithm.

49
00:01:33,160 --> 00:01:34,620
And this will be our first type of

50
00:01:34,720 --> 00:01:36,590
unsupervised learning, although there

51
00:01:36,790 --> 00:01:38,320
will be other types of unsupervised

52
00:01:39,020 --> 00:01:40,200
learning algorithms that we'll talk

53
00:01:40,320 --> 00:01:41,710
about later that finds other

54
00:01:42,130 --> 00:01:43,710
types of structure or other

55
00:01:43,920 --> 00:01:46,000
types of patterns in the data other than clusters.

56
00:01:46,900 --> 00:01:48,360
We'll talk about this afterwards, we will talk about clustering.

57
00:01:50,020 --> 00:01:51,210
So what is clustering good for?

58
00:01:51,380 --> 00:01:54,350
Early in this class I had already mentioned a few applications.

59
00:01:54,950 --> 00:01:56,540
One is market segmentation, where

60
00:01:56,670 --> 00:01:57,690
you may have a database of

61
00:01:57,770 --> 00:01:58,840
customers and want to group

62
00:01:59,070 --> 00:02:00,380
them into different market segments.

63
00:02:00,950 --> 00:02:02,590
So, you can sell to

64
00:02:02,720 --> 00:02:05,570
them separately or serve your different market segments better.

65
00:02:06,730 --> 00:02:08,370
Social network analysis, there are

66
00:02:08,580 --> 00:02:10,090
actually, you know, groups that have done

67
00:02:10,320 --> 00:02:12,590
this, things like looking at a

68
00:02:12,730 --> 00:02:14,540
group of people, social networks,

69
00:02:15,070 --> 00:02:16,390
so things like Facebook, Google plus

70
00:02:16,710 --> 00:02:18,260
or maybe information about who

71
00:02:18,430 --> 00:02:19,710
are the people that you

72
00:02:20,030 --> 00:02:21,110
email the most frequently and who

73
00:02:21,230 --> 00:02:22,170
are the people that they email

74
00:02:22,310 --> 00:02:23,600
the most frequently, and

75
00:02:23,750 --> 00:02:25,400
to find coherent groups of people.

76
00:02:26,500 --> 00:02:27,600
So, this would be another maybe

77
00:02:28,180 --> 00:02:28,850
clustering algorithm where, you know, you'd want

78
00:02:29,080 --> 00:02:32,200
to find who other coherent groups of friends in a social network.

79
00:02:33,140 --> 00:02:33,990
Here's something that one of my

80
00:02:34,140 --> 00:02:35,170
friends actually worked on, which is,

81
00:02:35,920 --> 00:02:37,200
use clustering to organize compute

82
00:02:37,670 --> 00:02:39,220
clusters or to organize data

83
00:02:39,440 --> 00:02:40,600
centers better because, if you

84
00:02:40,800 --> 00:02:42,450
know which computers in the

85
00:02:42,520 --> 00:02:44,990
data center are in the cluster there tend to work together.

86
00:02:45,400 --> 00:02:46,270
You can use that to reorganize

87
00:02:46,950 --> 00:02:48,390
your resources and how you

88
00:02:48,570 --> 00:02:50,120
lay out its network and

89
00:02:50,260 --> 00:02:52,040
how design your data center and communications.

90
00:02:53,140 --> 00:02:54,540
And lastly something that, actually

91
00:02:54,850 --> 00:02:55,910
another thing I worked on, using

92
00:02:56,130 --> 00:02:57,810
clustering algorithms to understand

93
00:02:58,400 --> 00:03:00,030
galaxy formation and using

94
00:03:00,280 --> 00:03:02,260
that to understand how, to

95
00:03:02,600 --> 00:03:03,860
understand astronomical detail.

96
00:03:06,550 --> 00:03:08,580
So, that's clustering which

97
00:03:08,890 --> 00:03:10,450
is our first example of

98
00:03:10,530 --> 00:03:12,650
an unsupervised learning algorithm.

99
00:03:13,090 --> 00:03:14,200
In the next video, we'll start to

100
00:03:14,370 --> 00:03:16,250
talk about a specific clustering algorithm.