﻿1
00:00:00,433 --> 00:00:02,100
Welcome back to your machine.

2
00:00:02,100 --> 00:00:06,600
Learning A to Z course is super excited
to have you back on board.

3
00:00:06,600 --> 00:00:10,866
And today we're kicking off the support
vector regression intuition.

4
00:00:11,066 --> 00:00:12,766
Super pumped about these tutorials.

5
00:00:12,766 --> 00:00:15,766
Got some very exciting slides
coming up for you.

6
00:00:15,833 --> 00:00:17,733
So what is support vector regression?

7
00:00:17,733 --> 00:00:23,333
Well, support vector regression was 
invented back in the 90s, by Vladimir

8
00:00:23,333 --> 00:00:28,066
Vapnik and his colleagues who were working
at the Bell Labs at the time.

9
00:00:28,066 --> 00:00:32,333
That was that AT&T Bell Labs down there
and knocked Nokia Bell Labs

10
00:00:33,000 --> 00:00:36,600
and, a lot of, support vector

11
00:00:36,600 --> 00:00:40,300
machine and support vector
regression are discussed in, Vladimir.

12
00:00:40,366 --> 00:00:44,333
Up next book,
the nature of Statistical Learning, 1992.

13
00:00:44,766 --> 00:00:48,000
In this course,
we will be covering both a support vector

14
00:00:48,366 --> 00:00:51,800
machine, in the part of the course
to do of classification

15
00:00:51,800 --> 00:00:53,133
and support vector regression.

16
00:00:53,133 --> 00:00:54,600
We'll be talking about it here.

17
00:00:54,600 --> 00:00:58,233
And in addition to that,
we'll also be talking about,

18
00:00:58,233 --> 00:01:02,100
kernel support vector machine and kernel,
a support

19
00:01:02,100 --> 00:01:06,400
vector regression and the kernel trick
and many other things like that.

20
00:01:06,600 --> 00:01:09,066
So there's a lot of exciting tutorials
on these topics,

21
00:01:09,066 --> 00:01:11,833
but for now we're just going
to limit ourselves to support vector

22
00:01:11,833 --> 00:01:15,633
regression, specifically linear
support vector regression.

23
00:01:16,000 --> 00:01:19,300
So here we go. here we've got two plots.

24
00:01:19,300 --> 00:01:22,333
And we need that in order to compare
SVR to,

25
00:01:22,700 --> 00:01:26,066
the simple linear regression
that'll really help us understand things.

26
00:01:26,266 --> 00:01:29,633
So here on the left
we've got some random plots, random dots.

27
00:01:29,933 --> 00:01:31,500
I'm going to copy them over to the right.

28
00:01:31,500 --> 00:01:33,200
So we know that these are identical.

29
00:01:33,200 --> 00:01:34,500
There's no tricks involved.

30
00:01:34,500 --> 00:01:36,300
There's an absolutely identical.

31
00:01:36,300 --> 00:01:39,300
this is an absolutely identical
set of data.

32
00:01:39,300 --> 00:01:41,166
And, 
let's start with the one on the left.

33
00:01:41,166 --> 00:01:44,133
We're going to apply a simple linear
regression.

34
00:01:44,133 --> 00:01:47,200
We've already discussed what it's like,
but let's quickly refresh.

35
00:01:47,200 --> 00:01:51,000
So basically we're going to 
have this line go through the data.

36
00:01:51,000 --> 00:01:52,766
And how is this line derived.

37
00:01:52,766 --> 00:01:55,800
Well a method called the ordinary least

38
00:01:55,833 --> 00:01:59,300
squares
method will be applied to find this line.

39
00:01:59,300 --> 00:02:03,933
Basically we want to minimize,
the distance between the

40
00:02:03,933 --> 00:02:07,800
this value
y, the actual value in the data

41
00:02:07,800 --> 00:02:10,333
and y hat basically
what it would have been on the trend line.

42
00:02:10,333 --> 00:02:12,466
We take the difference here
or difference here.

43
00:02:12,466 --> 00:02:14,233
We square it and we want to minimize that.

44
00:02:14,233 --> 00:02:16,166
That's ordinary least squares method.

45
00:02:16,166 --> 00:02:18,800
If essentially what we're doing
is minimize

46
00:02:18,800 --> 00:02:23,533
the error, we want to have a line
with the minimum error, possible.

47
00:02:23,533 --> 00:02:26,800
So that's that's the, intuition behind

48
00:02:27,200 --> 00:02:30,000
a simple linear regression,
something we already talked about.

49
00:02:30,000 --> 00:02:34,366
Now, how does this, support
vector regression work?

50
00:02:34,466 --> 00:02:35,866
Well, let's have a look on the right

51
00:02:35,866 --> 00:02:39,766
with SVR instead of a simple line,
you'll see a tube.

52
00:02:40,333 --> 00:02:43,566
And here you have
the regression line in the middle.

53
00:02:43,566 --> 00:02:46,266
But then there's this tube around.
And what does this tube do.

54
00:02:46,266 --> 00:02:50,800
Well this tube has a width of epsilon
and the width is measured vertically.

55
00:02:50,833 --> 00:02:54,766
This important along this axis not
perpendicular to the tube but vertically.

56
00:02:55,300 --> 00:03:00,166
And this tube itself is called
the epsilon insensitive tube.

57
00:03:00,300 --> 00:03:01,366
And what does that mean.

58
00:03:01,366 --> 00:03:05,800
Well,
that means that any, points in our data

59
00:03:05,800 --> 00:03:09,333
set that fall inside the tube,
they won't be.

60
00:03:09,333 --> 00:03:10,966
We'll be disregarding the error.

61
00:03:10,966 --> 00:03:14,933
So basically this tube,
think of it as a a margin of error

62
00:03:14,933 --> 00:03:18,366
that we are allowing our model to have

63
00:03:18,733 --> 00:03:21,966
and not care about any error inside here.

64
00:03:21,966 --> 00:03:26,666
So any discrepancy between
or any like distance between this,

65
00:03:26,966 --> 00:03:31,333
point over here and the line as in like
for instance here we could see

66
00:03:31,733 --> 00:03:34,733
let's look at 
which point is that that's,

67
00:03:35,300 --> 00:03:38,533
one, two, three, this third point, you
can see the line is even different, right?

68
00:03:38,533 --> 00:03:40,666
The results can be different
and probably will be different.

69
00:03:40,666 --> 00:03:43,633
So this third point here,
there's a distance between the line here.

70
00:03:43,633 --> 00:03:45,300
And we care about this error here.

71
00:03:45,300 --> 00:03:48,800
We don't care about this error
because it falls within this epsilon

72
00:03:49,000 --> 00:03:50,100
insensitive tube.

73
00:03:50,100 --> 00:03:52,900
So we're disregarding
any kind of errors in here.

74
00:03:52,900 --> 00:03:55,200
And that's kind of the key behind
support vector regression.

75
00:03:55,200 --> 00:03:58,966
And it gives a little bit of movement

76
00:03:58,966 --> 00:04:03,100
or a little bit of buffer to our model.

77
00:04:03,633 --> 00:04:06,900
And at the same time we have points
that are outside

78
00:04:06,900 --> 00:04:09,133
the epsilon insensitive tube.

79
00:04:09,133 --> 00:04:10,500
there they are.

80
00:04:10,500 --> 00:04:13,166
And for them we do care about the error.

81
00:04:13,166 --> 00:04:17,033
And here will be measured
as the distance between the, the point

82
00:04:17,033 --> 00:04:18,066
and the tube itself.

83
00:04:18,066 --> 00:04:20,400
So not the trend line,
but the tube itself.

84
00:04:20,400 --> 00:04:22,500
these distances have names.

85
00:04:22,500 --> 00:04:24,533
They're either C star.

86
00:04:24,533 --> 00:04:28,166
If, the point is below the tube, or see

87
00:04:28,333 --> 00:04:31,733
if the point is above the tube and,
they're called.

88
00:04:31,900 --> 00:04:34,533
So these values are called
slack variables.

89
00:04:34,533 --> 00:04:38,666
So it's either C star if it's below
C farthest star or if it's above.

90
00:04:39,000 --> 00:04:40,733
And we do care about the error.

91
00:04:40,733 --> 00:04:43,800
So we care about these distances
and the way we care about it.

92
00:04:43,800 --> 00:04:47,433
So we're going to try to avoid formulas
that will give additional

93
00:04:47,433 --> 00:04:51,233
reading something you can look into
further down at the end of this tutorial.

94
00:04:51,233 --> 00:04:55,866
But just for for completeness sake,
here is the formula.

95
00:04:55,966 --> 00:04:59,233
So in the case of OLS, it was a simple,

96
00:04:59,566 --> 00:05:02,933
ordinary least squared like that
here it's a bit more complex.

97
00:05:03,433 --> 00:05:05,700
we're not going to talk about this part
over here.

98
00:05:05,700 --> 00:05:08,000
but what we're focusing on is this.

99
00:05:08,000 --> 00:05:10,100
And we can see that we're minimizing.

100
00:05:10,100 --> 00:05:14,233
We want these, distances,
the sum of the sum of these distances

101
00:05:14,233 --> 00:05:16,933
to be minimal,
once again, will be additional

102
00:05:16,933 --> 00:05:18,733
reading at the end
if you'd like to go into it.

103
00:05:18,733 --> 00:05:21,700
But effectively, these are points.

104
00:05:21,700 --> 00:05:24,466
The ones that are outside
our tube are dictated

105
00:05:24,466 --> 00:05:28,133
what the tube will look like,
how the tube will be positioned

106
00:05:28,133 --> 00:05:31,233
so the error within
the tube is completely disregarded.

107
00:05:31,233 --> 00:05:34,200
We don't care about the error
unlike in the ordinary squares,

108
00:05:34,200 --> 00:05:37,733
so we're giving some kind of buffer
or flexibility to our tube.

109
00:05:38,066 --> 00:05:42,600
or like allowing it to, accounting

110
00:05:42,600 --> 00:05:47,100
for some kind of error that,
we might expect in the data.

111
00:05:47,100 --> 00:05:47,800
It's normal

112
00:05:47,800 --> 00:05:51,900
sometimes for there to be error,
but these ones, they are important to us.

113
00:05:52,333 --> 00:05:55,300
And, also one final thing.

114
00:05:55,300 --> 00:05:58,200
Why is this a method called support vector
regression?

115
00:05:58,200 --> 00:06:02,800
Well, because effectively these points,
all of these points outside.

116
00:06:03,000 --> 00:06:07,200
But at any point, actually any point on
this plot is a vector, right.

117
00:06:07,200 --> 00:06:07,800
Can be represented

118
00:06:07,800 --> 00:06:10,933
as a vector in this two dimensional space
or a multi dimensional space.

119
00:06:10,933 --> 00:06:13,000
If you have more features.

120
00:06:13,000 --> 00:06:15,766
So in this case it's kind of presented
by a two dimensional vector.

121
00:06:15,766 --> 00:06:18,400
So they are all these points are vectors.

122
00:06:18,400 --> 00:06:19,800
But the ones that we've highlighted

123
00:06:19,800 --> 00:06:23,633
in red, the ones outside the tube,
they're the support vectors

124
00:06:23,633 --> 00:06:27,633
because they are dictating
how this tube is created.

125
00:06:27,633 --> 00:06:32,700
So basically they're supporting
the structure or formation of this tube.

126
00:06:32,700 --> 00:06:35,133
And that's
why they're called support vectors.

127
00:06:35,133 --> 00:06:37,200
And that's why
this is a support vector regression.

128
00:06:38,166 --> 00:06:38,933
And so there we go.

129
00:06:38,933 --> 00:06:40,200
That's what it's all about.

130
00:06:40,200 --> 00:06:43,800
that just important to remember
the epsilon insensitive tube and that,

131
00:06:43,800 --> 00:06:45,300
support vector regression

132
00:06:45,300 --> 00:06:49,000
just cares about the errors of anything
that's lying outside this tube.

133
00:06:49,666 --> 00:06:52,666
And, to finish off, as promised, here's
the additional reading.

134
00:06:52,900 --> 00:06:56,666
so if you'd like to learn a bit more,
have a look at chapter four Support Vector

135
00:06:56,666 --> 00:07:00,500
Regression in a book called Efficient
Learning Machines theories, concepts,

136
00:07:00,500 --> 00:07:05,400
and Applications for Engineers and System
Design by Marietta Ward and Rahul Khanna.

137
00:07:05,800 --> 00:07:09,766
and here's a link here,
where it's aggregated on this,

138
00:07:10,066 --> 00:07:14,066
portal for, published work.

139
00:07:14,533 --> 00:07:19,166
something you'll note that
it might be a little bit confusing here.

140
00:07:19,166 --> 00:07:21,000
They say these are potential support
vectors.

141
00:07:21,000 --> 00:07:23,133
They're referring only to the ones close.

142
00:07:23,133 --> 00:07:27,900
I've had a look at different literature,
and, the literature I prefer is

143
00:07:27,900 --> 00:07:31,866
the one that says in this respect
is the one that says that the support

144
00:07:31,866 --> 00:07:36,500
vectors are the ones that are outside
and they are, outside.

145
00:07:36,500 --> 00:07:39,633
Any, any, basically any point outside
the tube is a support vector.

146
00:07:39,633 --> 00:07:44,100
So that's how we talked
discussed it inside this tutorial.

147
00:07:44,100 --> 00:07:46,733
But have a look at this different
nomenclature.

148
00:07:46,733 --> 00:07:50,666
Maybe that'll be a good perspective
to have a different view.

149
00:07:50,666 --> 00:07:55,866
But overall the first couple of paragraphs
describing the whole problem, they're

150
00:07:56,066 --> 00:07:58,033
very well written.
I liked how they were written.

151
00:07:58,033 --> 00:08:01,600
And I think, can be a valuable addition
if you're looking for additional reading.

152
00:08:02,133 --> 00:08:04,500
So there we go. That's, support vector
regression.

153
00:08:04,500 --> 00:08:07,200
Hope you enjoyed this tutorial
and look forward to seeing you next time.

154
00:08:07,200 --> 00:08:09,066
Until then, happy analyzing.
