1
00:00:00,166 --> 00:00:00,966
So let's try.

2
00:00:00,966 --> 00:00:05,600
Let's now reselect this section
to create a new regressor.

3
00:00:06,000 --> 00:00:09,433
So our part is already imported
so I don't need to select that again.

4
00:00:09,800 --> 00:00:12,800
And let's create this new regressor.

5
00:00:13,033 --> 00:00:13,800
Okay. Done.

6
00:00:13,800 --> 00:00:14,833
Perfect.

7
00:00:14,833 --> 00:00:17,666
New regressor created all properly. Okay.

8
00:00:17,666 --> 00:00:20,933
And now let's perhaps
first visualize the results

9
00:00:20,933 --> 00:00:24,533
to see if our model is now correct
before getting to the final verdict.

10
00:00:25,066 --> 00:00:28,066
Because, you know,
we need to validate the model. Now.

11
00:00:29,066 --> 00:00:32,200
So I'm selecting this section
and let's see what we get.

12
00:00:32,233 --> 00:00:35,233
Let's cross our fingers.

13
00:00:35,233 --> 00:00:35,933
Here we go.

14
00:00:35,933 --> 00:00:38,500
I'm going to zoom on the plot.

15
00:00:38,500 --> 00:00:40,700
And now the first word I have to say

16
00:00:40,700 --> 00:00:43,900
here is trap or red flag.

17
00:00:44,100 --> 00:00:47,000
Right now
we're just in front of a new trap.

18
00:00:47,000 --> 00:00:48,166
The model is improved.

19
00:00:48,166 --> 00:00:48,866
Definitely.

20
00:00:48,866 --> 00:00:52,033
We can definitely see
that we have more than one splits here.

21
00:00:52,233 --> 00:00:53,700
For example, here, that's a split.

22
00:00:53,700 --> 00:00:55,766
That's another split here and here.

23
00:00:55,766 --> 00:00:58,933
So okay, we solved the problem
of the number of splits.

24
00:00:59,400 --> 00:01:03,000
But according to what Kirill explained
in the intuition tutorial,

25
00:01:03,300 --> 00:01:07,000
do you think that the real shape
of a decision tree regression model.

26
00:01:07,600 --> 00:01:11,066
Well, because you know, what Kirill
explains is the algorithm of decision tree

27
00:01:11,066 --> 00:01:14,700
regression is that by considering
the entropy and the information gain,

28
00:01:14,933 --> 00:01:19,033
it's splitting the independent variables
into several intervals.

29
00:01:19,266 --> 00:01:22,200
So in the intuition tutorial
you had two independent variables.

30
00:01:22,200 --> 00:01:25,233
So the different intervals
formed some rectangles in which

31
00:01:25,233 --> 00:01:28,233
you took the average
of the dependent variable values.

32
00:01:28,300 --> 00:01:31,900
But here, since we are in one dimension,
that means that the algorithm

33
00:01:31,900 --> 00:01:34,900
will only take intervals
here of the independent variable.

34
00:01:34,933 --> 00:01:37,333
For example, that should be one interval.

35
00:01:37,333 --> 00:01:39,800
And here it looks like
it's the second one.

36
00:01:39,800 --> 00:01:41,266
And here it's the third one.

37
00:01:41,266 --> 00:01:43,166
And here a fourth one.

38
00:01:43,166 --> 00:01:43,566
Okay.

39
00:01:43,566 --> 00:01:46,666
So basically
it looks like we have four conditions

40
00:01:46,666 --> 00:01:49,600
and four intervals
that are making the splits.

41
00:01:49,600 --> 00:01:50,533
But as you understood

42
00:01:50,533 --> 00:01:54,566
in the intuition tutorial
it's taking the average in each interval.

43
00:01:54,900 --> 00:01:58,500
So if it's taking the average
how do you want to have this

44
00:01:58,500 --> 00:02:00,633
straight line here that is not horizontal.

45
00:02:00,633 --> 00:02:02,366
Because you know what the decision tree

46
00:02:02,366 --> 00:02:05,766
regression is doing
is that in each interval it's calculating

47
00:02:05,766 --> 00:02:08,766
the average
of the dependent variable salaries.

48
00:02:08,800 --> 00:02:11,800
And therefore for all the levels
contained in this interval,

49
00:02:12,133 --> 00:02:15,100
the value of the prediction
should be a constant equal

50
00:02:15,100 --> 00:02:18,100
to this average of the dependent variable
in this interval.

51
00:02:18,333 --> 00:02:20,600
And here, as we can see,
it's not a constant.

52
00:02:20,600 --> 00:02:23,266
You know the prediction here
is not the same as the prediction here.

53
00:02:23,266 --> 00:02:27,000
So either
it's considering an infinity of intervals

54
00:02:27,000 --> 00:02:30,433
with different constants
in each of those infinite intervals.

55
00:02:30,800 --> 00:02:33,233
Or either we have a problem here.

56
00:02:33,233 --> 00:02:34,766
Of course it's not the first option.

57
00:02:34,766 --> 00:02:37,800
Of course, the decision tree regression
is not considering

58
00:02:37,800 --> 00:02:41,000
an infinity of intervals
between this level and this level.

59
00:02:41,500 --> 00:02:43,700
So it's definitely the second option.

60
00:02:43,700 --> 00:02:46,266
So now
do you see where the problem comes from.

61
00:02:46,266 --> 00:02:49,500
Well the answer is in our regression
template.

62
00:02:49,966 --> 00:02:53,966
Because what we observe here
is only due to the resolution

63
00:02:53,966 --> 00:02:57,366
we picked to plot these decision tree
regression results,

64
00:02:57,900 --> 00:03:00,533
because we are actually plotting
the predictions

65
00:03:00,533 --> 00:03:03,566
for each of the ten levels
incremented by one.

66
00:03:03,600 --> 00:03:05,700
That means that here,
you know, it's only plotting

67
00:03:05,700 --> 00:03:09,000
the predictions of the ten salaries
corresponding to the ten levels,

68
00:03:09,333 --> 00:03:12,366
and then it's joining the predictions
by a straight line here,

69
00:03:12,533 --> 00:03:16,300
because it had no predictions
to plot in this interval

70
00:03:16,300 --> 00:03:18,866
here of the independent variable level.

71
00:03:18,866 --> 00:03:22,900
And the fact that it's a problem
for this new non-linear regression model,

72
00:03:22,900 --> 00:03:26,500
the decision tree
model is due to a very specific reason

73
00:03:26,966 --> 00:03:28,800
for the previous non-linear
regression models.

74
00:03:28,800 --> 00:03:31,700
We could use the code
that generated this plot

75
00:03:31,700 --> 00:03:33,933
because the models
were actually continuous.

76
00:03:33,933 --> 00:03:36,700
So for example
in the polynomial regression model

77
00:03:36,700 --> 00:03:39,666
well between these prediction
and these prediction,

78
00:03:39,666 --> 00:03:42,400
well it was actually almost a straight
line here.

79
00:03:42,400 --> 00:03:45,633
However, right now we are facing
a new kind of regression model.

80
00:03:46,000 --> 00:03:47,766
Remember the first kind of regression
model

81
00:03:47,766 --> 00:03:50,400
we studied
was the linear regression model.

82
00:03:50,400 --> 00:03:54,333
Then the second kind of regression model
we saw was the nonlinear regression model.

83
00:03:54,633 --> 00:03:58,000
And now we're facing
a new kind of regression model.

84
00:03:58,200 --> 00:04:02,100
It's the nonlinear
and non continuous regression model.

85
00:04:03,000 --> 00:04:05,633
Indeed all the previous regression models
that we saw,

86
00:04:05,633 --> 00:04:09,600
whether they were linear or not linear
they were all continuous.

87
00:04:09,866 --> 00:04:13,433
But here the decision tree regression
model is not continuous.

88
00:04:13,433 --> 00:04:14,833
And this is the first non

89
00:04:14,833 --> 00:04:17,833
continuous machine learning model
we are seeing together.

90
00:04:17,966 --> 00:04:18,666
And so do you know

91
00:04:18,666 --> 00:04:21,900
what is the best way to visualize
a non continuous regression model.

92
00:04:21,966 --> 00:04:24,333
Well as I was telling you the answer

93
00:04:24,333 --> 00:04:27,433
the solution for
this is in our regression template.

94
00:04:27,766 --> 00:04:29,133
So let's have a look.

95
00:04:29,133 --> 00:04:32,733
And actually
we need to take the code section

96
00:04:32,733 --> 00:04:36,633
that visualize the regression
model results in higher resolution.

97
00:04:37,100 --> 00:04:41,033
So let's take this
and let's actually go back to our decision

98
00:04:41,033 --> 00:04:45,166
tree regression file
and replace this code here.

99
00:04:45,166 --> 00:04:47,233
Because this is totally not appropriate

100
00:04:47,233 --> 00:04:50,533
for our decision tree regression model
because it's a non continuous model.

101
00:04:50,866 --> 00:04:55,433
And so we need to replace this code
by the same but for the high resolution.