1
00:00:00,233 --> 00:00:03,866
Hello my friends,
and welcome to this new practical activity

2
00:00:03,900 --> 00:00:07,200
on this time support vector regression.

3
00:00:07,466 --> 00:00:11,866
So I have to start by telling you
that the model we're about to implement

4
00:00:11,866 --> 00:00:15,633
will be slightly more advanced
than the models were built before,

5
00:00:15,633 --> 00:00:16,800
but it's totally fine.

6
00:00:16,800 --> 00:00:19,900
We will do it together
and we will succeed 100%.

7
00:00:20,366 --> 00:00:22,833
So why is it slightly more advanced?

8
00:00:22,833 --> 00:00:27,000
Well, that's because we will have to play
a lot with feature scaling.

9
00:00:27,000 --> 00:00:30,733
And I can tell you that after
this implementation of the SVR model,

10
00:00:30,866 --> 00:00:34,833
you will be a master in feature scaling
because you will know

11
00:00:34,866 --> 00:00:38,000
not only how to apply the feature
scaling transformation,

12
00:00:38,166 --> 00:00:41,100
but also how to apply the inverse
transformation.

13
00:00:41,100 --> 00:00:43,900
You know,
to go back to the original scaling.

14
00:00:43,900 --> 00:00:46,900
So you will know all the Sep tools,
let's say

15
00:00:46,900 --> 00:00:49,900
a feature scaling,
and you will handle them like a probe.

16
00:00:49,900 --> 00:00:50,600
All right.

17
00:00:50,600 --> 00:00:52,933
So if you're ready let's start.

18
00:00:52,933 --> 00:00:54,800
And just before we go inside

19
00:00:54,800 --> 00:00:58,500
the folder, let's make sure here
that we are all on the same page.

20
00:00:58,733 --> 00:01:02,566
I gave you the link to this folder
containing all the codes and data sets

21
00:01:02,766 --> 00:01:06,400
right before this tutorial in an oracle,
so make sure to connect to it.

22
00:01:06,600 --> 00:01:08,933
And now we should all be on the same page.

23
00:01:08,933 --> 00:01:12,633
And therefore we're going to go
of course to part two regression.

24
00:01:12,800 --> 00:01:17,200
And then of course to support vector
regression SVR.

25
00:01:17,766 --> 00:01:20,100
All right. And as usual
we're going to start with Python.

26
00:01:20,100 --> 00:01:23,700
So we're going to go to that Python
folder inside which will find

27
00:01:23,866 --> 00:01:26,166
the same data set as before.

28
00:01:26,166 --> 00:01:30,500
Because you know I would like to compare
the performance of different regression

29
00:01:30,500 --> 00:01:34,533
models on this data set,
which proved to have some non-linear

30
00:01:34,533 --> 00:01:37,200
relationships,
you know, between the feature,

31
00:01:37,200 --> 00:01:40,433
which is the level you know,
the position level going from 1 to 10

32
00:01:40,666 --> 00:01:45,466
and the salary going from 45,000
to $1 million per year.

33
00:01:45,833 --> 00:01:48,766
So here again,
this is exactly the same scenario.

34
00:01:48,766 --> 00:01:53,933
We want to train this time
a support vector regression model to learn

35
00:01:53,933 --> 00:01:57,900
and understand the correlations between
these position levels and these salaries,

36
00:01:58,166 --> 00:02:02,400
and to quickly remind the context,
we are hiring a new person

37
00:02:02,700 --> 00:02:07,500
who is expecting 160 K salary,
justifying this by the fact

38
00:02:07,500 --> 00:02:11,900
that this person earned a 160 K salary
in the previous company.

39
00:02:12,333 --> 00:02:15,100
This data is exactly the data
of the previous company.

40
00:02:15,100 --> 00:02:16,700
You know, with the different position

41
00:02:16,700 --> 00:02:20,400
levels from business analyst
to CEO and their corresponding salary.

42
00:02:20,633 --> 00:02:24,800
And so not only we want to train a model
to learn these relationships,

43
00:02:24,966 --> 00:02:27,966
but also we want to deploy this model
to predict

44
00:02:27,966 --> 00:02:30,966
the salary that this person
had in this previous company,

45
00:02:31,166 --> 00:02:35,700
knowing that indeed, this person
was a region manager for a couple of years

46
00:02:35,700 --> 00:02:40,000
and therefore is considered
to have a position level of 6.5.

47
00:02:40,200 --> 00:02:43,066
So that's the exact same scenario, exact
same data set.

48
00:02:43,066 --> 00:02:47,700
And so now let's build the support vector
regression model on this data

49
00:02:47,700 --> 00:02:49,533
set to see if it performs

50
00:02:49,533 --> 00:02:53,100
better than the previous model
which showed great results.

51
00:02:53,366 --> 00:02:55,333
The polynomial regression model.

52
00:02:55,333 --> 00:02:55,800
All right.

53
00:02:55,800 --> 00:02:56,633
So let's do this.

54
00:02:56,633 --> 00:03:01,266
Let's close
this and let's open our support vector

55
00:03:01,266 --> 00:03:05,033
regression implementation
either with Google Colaboratory

56
00:03:05,033 --> 00:03:10,366
if you love it, or with Jupyter Notebook
because indeed it is an IP and B file.

57
00:03:10,600 --> 00:03:11,400
All right.

58
00:03:11,400 --> 00:03:13,033
So that's the whole implementation.

59
00:03:13,033 --> 00:03:16,200
Let me quickly show you
the structure of the implementation.

60
00:03:16,400 --> 00:03:19,266
We're going to start
by importing the libraries as usual

61
00:03:19,266 --> 00:03:22,766
then importing the data set
then applying feature scaling.

62
00:03:22,766 --> 00:03:23,733
So this is interesting.

63
00:03:23,733 --> 00:03:27,500
This time we have to apply feature scaling
because in the SVR model

64
00:03:27,500 --> 00:03:30,600
there is not this,
you know, explicit equation

65
00:03:30,866 --> 00:03:33,866
of the dependent variable
with respect to the features.

66
00:03:33,966 --> 00:03:37,833
And mostly there are not those,
you know, coefficients multiplying

67
00:03:37,833 --> 00:03:39,000
each of the features

68
00:03:39,000 --> 00:03:43,166
and therefore not compensating
with lower values for the features.

69
00:03:43,166 --> 00:03:44,833
Taking high values. No.

70
00:03:44,833 --> 00:03:50,133
This time the support vector regression
model has an implicit equation

71
00:03:50,133 --> 00:03:53,600
of you know, the dependent variable
with respect to the features.

72
00:03:53,800 --> 00:03:57,466
So we don't have such coefficients,
and we will definitely

73
00:03:57,733 --> 00:04:00,600
have to apply feature
scaling for this model.

74
00:04:00,600 --> 00:04:04,633
So you see you start to understand when
and when not to apply features.

75
00:04:04,633 --> 00:04:05,600
Getting well.

76
00:04:05,600 --> 00:04:09,500
You know we don't have to apply feature
scaling for linear regression models

77
00:04:09,500 --> 00:04:10,300
where you have indeed

78
00:04:10,300 --> 00:04:14,000
those coefficients that can compensate
with the high values of the features.

79
00:04:14,333 --> 00:04:17,933
And so of course these include
simple linear regression, multiple linear

80
00:04:17,933 --> 00:04:19,933
regression and polynomial regression.

81
00:04:19,933 --> 00:04:23,333
And you will see later on in the course
many other models.

82
00:04:23,633 --> 00:04:28,600
And for other models
which usually have an implicit equation,

83
00:04:28,600 --> 00:04:30,866
you know, an implicit relationship

84
00:04:30,866 --> 00:04:34,800
between the dependent variable y
and the features x.

85
00:04:34,933 --> 00:04:38,700
Well, usually for these models
we have to apply feature scaling.

86
00:04:38,700 --> 00:04:39,433
All right.

87
00:04:39,433 --> 00:04:42,866
Then we will of course
trained SVM model on the whole data set.

88
00:04:42,866 --> 00:04:44,233
You know this time is the same.

89
00:04:44,233 --> 00:04:48,233
We want split the whole data
set into a training set and a test set,

90
00:04:48,433 --> 00:04:51,800
because we want to leverage
the maximum data to learn

91
00:04:51,800 --> 00:04:55,466
these correlations between
those position levels and the salaries.

92
00:04:55,633 --> 00:04:56,833
So we won't do that split.

93
00:04:56,833 --> 00:05:00,400
And we will directly train
the SVR model on the whole data set.

94
00:05:00,866 --> 00:05:04,333
Then after this training, well,
we'll have a smart SVR model,

95
00:05:04,500 --> 00:05:08,433
which therefore we're going to use
to predict this new result and exactly

96
00:05:08,700 --> 00:05:12,300
that salary of this position level of 6.5.

97
00:05:12,500 --> 00:05:15,633
And we will compare, of course,
that prediction with the prediction

98
00:05:15,633 --> 00:05:18,733
of the polynomial regression model,
which I've kept here.

99
00:05:19,333 --> 00:05:23,700
And then of course, we will visualize
SVR results in low resolution

100
00:05:23,700 --> 00:05:28,733
and high resolution to see, of course,
the regression curve of the SVR model.

101
00:05:29,200 --> 00:05:29,666
All right.

102
00:05:29,666 --> 00:05:33,633
I'm sure you noticed that
I did not click on each of these contents.

103
00:05:33,633 --> 00:05:34,833
Well, I did that on purpose.

104
00:05:34,833 --> 00:05:37,666
It's because I don't want to reveal the 
prediction.

105
00:05:37,666 --> 00:05:40,500
You know, the predicted salary
by the SVR model.

106
00:05:40,500 --> 00:05:44,533
I want us to save the surprise until,
you know, we execute that cell

107
00:05:44,533 --> 00:05:46,500
to return that predicted salary.