﻿1
00:00:00,066 --> 00:00:00,733
Okay.

2
00:00:00,733 --> 00:00:04,866
Actually, this is all we had to change
in this regression template.

3
00:00:05,166 --> 00:00:08,100
And now we are ready
to execute the sections one by one

4
00:00:08,100 --> 00:00:11,133
to create our model
and find out about the final results

5
00:00:11,133 --> 00:00:14,133
and the final verdict,
whether it's truth or bluff.

6
00:00:14,200 --> 00:00:16,500
So let's do it.

7
00:00:16,500 --> 00:00:19,500
I'm going to import the dataset first.

8
00:00:19,500 --> 00:00:20,433
All right here we go.

9
00:00:20,433 --> 00:00:22,200
Well import it. We can have a look.

10
00:00:22,200 --> 00:00:25,366
So these are our data set with the levels
as the independent variable

11
00:00:25,366 --> 00:00:28,366
and the salaries
as the dependent variable.

12
00:00:28,666 --> 00:00:29,700
Okay great.

13
00:00:29,700 --> 00:00:33,766
And now we need to split the dataset
into the training set and the test set.

14
00:00:33,800 --> 00:00:34,800
No need for feature scaling

15
00:00:34,800 --> 00:00:38,700
because we're using the very popular
E10 71 library that includes it.

16
00:00:39,233 --> 00:00:41,400
And now let's create our model.

17
00:00:41,400 --> 00:00:44,233
So I'm going to select
this whole section here.

18
00:00:44,233 --> 00:00:48,600
Press Command and Control Plus
and do the cute done regress or created.

19
00:00:48,900 --> 00:00:49,800
Great.

20
00:00:49,800 --> 00:00:52,666
And now let's predict the new results
which I'd like to start

21
00:00:52,666 --> 00:00:56,500
by visualizing the SVR results
first or predicting the new result.

22
00:00:56,500 --> 00:01:00,833
Well, let's actually predict the result
because this is the most exciting step.

23
00:01:01,000 --> 00:01:02,833
And let's keep the best for the end.

24
00:01:02,833 --> 00:01:04,533
But this step is actually very exciting.

25
00:01:04,533 --> 00:01:05,700
Also, because

26
00:01:05,700 --> 00:01:09,300
we are getting the final prediction,
we are getting the final predicted salary.

27
00:01:09,300 --> 00:01:12,600
So you know it says you want
if you're doing actually this code

28
00:01:12,600 --> 00:01:15,633
by yourself before me,
that's actually a very good exercise.

29
00:01:15,633 --> 00:01:17,533
So I really encourage you to do this.

30
00:01:17,533 --> 00:01:20,300
Choose whatever you want to execute first.

31
00:01:20,300 --> 00:01:21,300
It's ready anyway.

32
00:01:21,300 --> 00:01:23,300
You don't need to change anything
more now.

33
00:01:23,300 --> 00:01:26,100
So I'm going to execute that right now.

34
00:01:26,100 --> 00:01:29,100
And actually we obtain a predicted salary

35
00:01:29,266 --> 00:01:31,933
of $177,000.

36
00:01:31,933 --> 00:01:33,100
Great.

37
00:01:33,100 --> 00:01:36,600
So remember this employee who we are
negotiating the future salary right

38
00:01:36,600 --> 00:01:39,733
now, said that its previous
salary was 160 K.

39
00:01:40,033 --> 00:01:45,166
And our model predicted
that its previous salary was 177 K.

40
00:01:45,500 --> 00:01:49,200
So first of all, that's actually close
to what this employee said.

41
00:01:49,200 --> 00:01:52,133
And besides,
it's on the good side of the negotiation.

42
00:01:52,133 --> 00:01:53,600
So that's actually pretty good.

43
00:01:53,600 --> 00:01:57,333
And we can be satisfied with this result
and our model.

44
00:01:57,333 --> 00:02:01,866
But to be really satisfied let's see
what's happening with the graphic result.

45
00:02:01,866 --> 00:02:04,866
So I'm going to execute this section

46
00:02:04,866 --> 00:02:07,733
and let's have a look at the SVR model.

47
00:02:09,600 --> 00:02:10,933
And here it is.

48
00:02:10,933 --> 00:02:14,233
I'm going to zoom on it and better okay.

49
00:02:14,700 --> 00:02:18,900
So first of all this model fits
very well most of the data points.

50
00:02:19,266 --> 00:02:22,266
As a reminder the real observation
points are the red points here.

51
00:02:22,500 --> 00:02:26,966
And all the points in this blue curve here
which is the SVR model itself

52
00:02:27,233 --> 00:02:28,466
are the prediction points.

53
00:02:28,466 --> 00:02:32,100
So for example, if we're taking this
red observation point here,

54
00:02:32,366 --> 00:02:35,366
well the prediction is perfect
because the prediction point

55
00:02:35,366 --> 00:02:38,333
is the projection of this red points
on the blue curve.

56
00:02:38,333 --> 00:02:41,000
So it's actually the red point itself
because the red point

57
00:02:41,000 --> 00:02:42,466
is on the blue curve.

58
00:02:42,466 --> 00:02:43,866
So that makes a perfect prediction.

59
00:02:43,866 --> 00:02:45,133
But here for example

60
00:02:45,133 --> 00:02:48,133
if we take a less accurate prediction
but still a very good one,

61
00:02:48,200 --> 00:02:50,100
then we take the real observation
point here.

62
00:02:50,100 --> 00:02:53,100
The red point
we projected on the blue curve.

63
00:02:53,166 --> 00:02:57,000
And that's the difference
between the real salary

64
00:02:57,000 --> 00:02:58,866
and the predicted salary, which is here.

65
00:02:58,866 --> 00:03:02,533
If we projected back on the y axis,
which contains the salaries.

66
00:03:02,700 --> 00:03:07,100
But you can see that for all of these
points from this one to actually this one,

67
00:03:07,433 --> 00:03:10,966
the blue curve is actually getting
very close to the real observation

68
00:03:10,966 --> 00:03:12,533
points, the red points.

69
00:03:12,533 --> 00:03:15,533
And so the predictions
are very close to real results.

70
00:03:15,866 --> 00:03:21,100
But that's for all these points
of the data set except for this one here.

71
00:03:21,100 --> 00:03:25,200
This one is left alone
which by the way corresponds to the CEO.

72
00:03:25,200 --> 00:03:26,733
So I'm sorry for the CEO,

73
00:03:26,733 --> 00:03:31,500
but the reason for what is happening here
is that this is actually an outlier.

74
00:03:31,600 --> 00:03:33,266
Not to call the CEO, not liar.

75
00:03:33,266 --> 00:03:36,266
But this is an outlier
because as you can see,

76
00:03:36,466 --> 00:03:40,766
this point is actually far from
the other points in terms of salaries.

77
00:03:41,100 --> 00:03:45,500
The CEO has a much higher salary
than the previous positions.

78
00:03:45,500 --> 00:03:47,766
So for the SVR model, that's an outlier.

79
00:03:47,766 --> 00:03:49,800
And therefore he actually didn't
consider it.

80
00:03:49,800 --> 00:03:53,100
It's like it excluded this point
in the model by not looking at it

81
00:03:53,400 --> 00:03:56,733
and making its predictions
on these points here.

82
00:03:57,266 --> 00:04:00,466
So that's specifically due to the SVR
algorithm itself.

83
00:04:00,466 --> 00:04:02,466
But there are a lot of parameters.

84
00:04:02,466 --> 00:04:05,533
And you can actually play
with the parameters to change

85
00:04:05,533 --> 00:04:08,833
the way the SVR model perceives outliers.

86
00:04:09,000 --> 00:04:10,033
So for example,

87
00:04:10,033 --> 00:04:13,900
these parameters are the penalty
parameters, the regularization parameters.

88
00:04:14,233 --> 00:04:16,766
It's most of the time
well described in the description.

89
00:04:16,766 --> 00:04:20,300
And of course the E 1071 library includes
such techniques.

90
00:04:21,000 --> 00:04:25,533
But we're not going to do it here in this
tutorial because we actually don't need

91
00:04:25,533 --> 00:04:29,800
to get a good prediction of the CEO salary
and remember what we actually needed.

92
00:04:29,800 --> 00:04:30,633
What is a good prediction

93
00:04:30,633 --> 00:04:33,633
of the previous salary of this employee
we are negotiating with?

94
00:04:33,833 --> 00:04:36,300
And this employee has a level 6.5.

95
00:04:36,300 --> 00:04:39,533
And around this point we can see that
the SVR model fits very well

96
00:04:39,533 --> 00:04:40,633
to our data set.

97
00:04:40,633 --> 00:04:45,366
And we actually get a prediction of 177
K quite close to the real

98
00:04:45,366 --> 00:04:49,533
or the mentioned salary of this employee,
which is 160 K.

99
00:04:49,900 --> 00:04:51,100
So that's actually pretty good.

100
00:04:51,100 --> 00:04:54,233
And therefore the verdict of truth

101
00:04:54,233 --> 00:04:57,866
or bluff
according to our SVR model is rather true.
