﻿1
00:00:00,300 --> 00:00:02,400
But now the most exciting part is coming

2
00:00:02,400 --> 00:00:05,400
because we're going to be visualizing
the training set results.

3
00:00:06,200 --> 00:00:06,900
And actually, right now

4
00:00:06,900 --> 00:00:10,800
we don't have to change anything
except maybe the title here classifier.

5
00:00:11,300 --> 00:00:13,966
And then here you need to specify kernel

6
00:00:13,966 --> 00:00:18,066
SVM, kernel SVM
as well as here in the test set.

7
00:00:18,066 --> 00:00:21,166
So here kernel SVM as well.

8
00:00:22,966 --> 00:00:23,366
All right.

9
00:00:23,366 --> 00:00:27,733
Now we can get some coffee and just
execute this and watch the results.

10
00:00:28,066 --> 00:00:30,000
So we're going to do this.

11
00:00:30,000 --> 00:00:32,966
We're going to first look at the results

12
00:00:32,966 --> 00:00:35,133
here and press Command and control.

13
00:00:35,133 --> 00:00:38,133
Press enter to execute.

14
00:00:40,666 --> 00:00:44,000
And wow these are the beautiful results
of kernel SVM.

15
00:00:44,533 --> 00:00:47,400
We can see how it's making this curve here
to classify.

16
00:00:47,400 --> 00:00:50,600
Well most of the users
in the social network here are since

17
00:00:50,600 --> 00:00:53,600
we have a two dimensional space
here for the data set,

18
00:00:53,600 --> 00:00:57,533
that means that the data set was mapped
to a three dimensional space,

19
00:00:58,000 --> 00:00:59,800
and then on this dimensional space,

20
00:00:59,800 --> 00:01:02,700
it managed to make the data
linearly separable,

21
00:01:02,700 --> 00:01:07,333
so that it can find a linear hyperplane
that separates in a much better way.

22
00:01:07,333 --> 00:01:08,766
The two classes.

23
00:01:08,766 --> 00:01:12,533
And then it projected back to this
two dimensional space

24
00:01:12,733 --> 00:01:18,066
to finally obtain this curve here
that separates in a much better way

25
00:01:18,066 --> 00:01:21,900
the two classes than the linear classifier
you would do with a straight line here,

26
00:01:22,100 --> 00:01:26,133
as we saw for logistic regression
or SVM, SVM with the linear kernel.

27
00:01:26,566 --> 00:01:27,833
So that's what happened.

28
00:01:27,833 --> 00:01:29,700
That's what happened behind the scenes.

29
00:01:29,700 --> 00:01:31,500
The kernel SVM did an amazing job.

30
00:01:31,500 --> 00:01:33,266
So we can say congrats.

31
00:01:33,266 --> 00:01:36,533
Okay, so for those of you
who are seeing this for the first time,

32
00:01:36,733 --> 00:01:38,566
I'll just quickly remind what it is.

33
00:01:38,566 --> 00:01:41,300
The points here are the true
observations, the true results.

34
00:01:41,300 --> 00:01:45,300
So the red points are the users
who didn't buy the SUV in reality,

35
00:01:45,533 --> 00:01:48,566
and the green points are
the users who bought the SUV.

36
00:01:48,800 --> 00:01:51,766
And then the regions that we see
here are the prediction regions.

37
00:01:51,766 --> 00:01:54,833
So the red region is the region
where our model predicts

38
00:01:54,833 --> 00:01:58,133
that the user doesn't buy the SUV
and the green region is the one

39
00:01:58,133 --> 00:02:01,133
where our model predicts
that the user buys the SUV.

40
00:02:01,166 --> 00:02:05,033
So for example, that means
that for this particular user red points,

41
00:02:05,266 --> 00:02:08,733
which is in the green region,
that makes an incorrect prediction

42
00:02:08,733 --> 00:02:12,133
because the truth is,
the user didn't buy the SUV.

43
00:02:12,466 --> 00:02:15,200
But since this user
is in the green region, that means

44
00:02:15,200 --> 00:02:18,500
that the model predicted
that this user bought the SUV,

45
00:02:18,733 --> 00:02:20,800
even if that's not
what happened in real life.

46
00:02:20,800 --> 00:02:22,800
So that's where the graph is about.

47
00:02:22,800 --> 00:02:27,600
And here is the nonlinear separator
of the kernel SVM algorithm.

48
00:02:27,900 --> 00:02:32,900
And thanks to this nonlinear separator
our kernel SVM could classify correctly

49
00:02:32,900 --> 00:02:36,000
those users here,
which wasn't the case for the linear

50
00:02:36,000 --> 00:02:38,300
separators
because since we had the straight lines,

51
00:02:38,300 --> 00:02:41,433
it couldn't
catch these users in the right category.

52
00:02:41,433 --> 00:02:43,666
That is the green category,
because those users

53
00:02:43,666 --> 00:02:45,566
actually bought the SUV in real life.

54
00:02:45,566 --> 00:02:47,166
So they should be in the green region.

55
00:02:47,166 --> 00:02:49,400
And this wasn't the case
for the linear separator.

56
00:02:49,400 --> 00:02:50,200
And now it's the case

57
00:02:50,200 --> 00:02:53,800
for the kernel SVM separator
because it's a nonlinear classifier.

58
00:02:54,500 --> 00:02:55,000
Perfect.

59
00:02:55,000 --> 00:02:59,533
Now let's have a look at the test results
to see how our model behaves

60
00:02:59,533 --> 00:03:02,366
on new observations,
to see if it can still classify

61
00:03:02,366 --> 00:03:05,166
and predict
the results of new observations.

62
00:03:05,166 --> 00:03:07,166
So we're going to do this. We only have to

63
00:03:08,700 --> 00:03:09,566
select this.

64
00:03:09,566 --> 00:03:10,533
Press command and control.

65
00:03:10,533 --> 00:03:13,533
Press enter to execute.

66
00:03:13,933 --> 00:03:16,433
And here are the test set results okay.

67
00:03:16,433 --> 00:03:18,200
So that looks great as well.

68
00:03:18,200 --> 00:03:21,600
First an important thing to understand
is that the regions that we see

69
00:03:21,600 --> 00:03:24,600
here are the same regions
as we saw in the data set.

70
00:03:24,700 --> 00:03:27,633
If you go back to the training set
and then go back to the test set,

71
00:03:27,633 --> 00:03:31,933
you can see that the regions here
don't change, only the observation

72
00:03:31,933 --> 00:03:34,933
points change because of course
these are new observation points.

73
00:03:35,100 --> 00:03:36,900
And so here we can see that

74
00:03:36,900 --> 00:03:40,933
the kernel SVM did an amazing job
at classifying new observations

75
00:03:40,933 --> 00:03:45,933
because all these test set observations
here are the users who didn't buy the SUV,

76
00:03:45,933 --> 00:03:48,933
because the points are red
and they are in the red region.

77
00:03:48,933 --> 00:03:52,500
That is the region where our classifier
predicts that the users don't buy the SUV.

78
00:03:52,800 --> 00:03:54,633
So that's correct. Predictions.

79
00:03:54,633 --> 00:03:55,900
Same for these guys here.

80
00:03:55,900 --> 00:03:58,966
All the green points
that fall in the green region.

81
00:03:59,333 --> 00:04:02,333
And of course we have
those incorrect predictions

82
00:04:02,433 --> 00:04:05,000
which are for example,
this green user here

83
00:04:05,000 --> 00:04:08,400
who is in the red region
as well as those two green here.

84
00:04:08,400 --> 00:04:09,533
So that's three.

85
00:04:09,533 --> 00:04:12,000
Then we have the red users. This one,
this one.

86
00:04:12,000 --> 00:04:17,166
So 456789 and actually ten.

87
00:04:17,200 --> 00:04:21,100
We can see that there is a little green
point here hiding behind the red points

88
00:04:22,033 --> 00:04:22,466
okay. Yeah.

89
00:04:22,466 --> 00:04:24,233
So that's definitely a very good job.

90
00:04:24,233 --> 00:04:26,133
You will see how the next classifiers

91
00:04:26,133 --> 00:04:29,700
will make a different style
of classification of separation.

92
00:04:30,133 --> 00:04:31,800
You will see that it won't be a curve.

93
00:04:31,800 --> 00:04:32,833
It will be something else.

94
00:04:32,833 --> 00:04:35,766
And I will let you find out
about the surprise.

95
00:04:35,766 --> 00:04:38,933
But definitely
the kernel SVM is a great classifier

96
00:04:38,933 --> 00:04:41,366
when your data is not linearly separable.

97
00:04:41,366 --> 00:04:45,033
And we can see how it definitely improved
the results of the linear classifiers

98
00:04:45,033 --> 00:04:48,033
that we saw
with logistic regression and SVM.

99
00:04:48,066 --> 00:04:48,466
All right.

100
00:04:48,466 --> 00:04:51,733
So thank you for watching this tutorial
about kernel SVM.

101
00:04:51,900 --> 00:04:53,033
I hope you enjoyed it.

102
00:04:53,033 --> 00:04:56,400
There will be a few more surprises
with the next classifiers, as I told you.

103
00:04:56,533 --> 00:04:58,200
Until then, enjoy machine learning.
