1
00:00:00,566 --> 00:00:00,933
All right.

2
00:00:00,933 --> 00:00:05,200
So now, my friends,
we're going to rerun all this, right?

3
00:00:05,200 --> 00:00:07,133
Because we're basically done.

4
00:00:07,133 --> 00:00:11,000
And so to run all this again
we're going to first upload

5
00:00:11,000 --> 00:00:14,466
the data set because it is not yet
uploaded in the notebook.

6
00:00:14,900 --> 00:00:17,866
So upload and then make sure to find

7
00:00:17,866 --> 00:00:21,366
your whole machine learning is that folder
containing all the codes and data sets.

8
00:00:21,633 --> 00:00:23,033
So then we're going to go inside.

9
00:00:23,033 --> 00:00:25,600
Then we're going to go to part
three classification.

10
00:00:25,600 --> 00:00:29,866
And then K nearest neighbors Kiernan
and then Python.

11
00:00:29,866 --> 00:00:35,066
And then we select our social network
add data set click open.

12
00:00:35,566 --> 00:00:38,300
Now it's going to be uploaded
on the notebook.

13
00:00:38,300 --> 00:00:39,933
Perfect. There we go.

14
00:00:39,933 --> 00:00:41,533
Now we can run everything.

15
00:00:41,533 --> 00:00:45,000
And to do this
well we're going to click runtime here.

16
00:00:45,000 --> 00:00:50,566
And then simply run
and let's just start from here okay.

17
00:00:50,566 --> 00:00:51,433
Great.

18
00:00:51,433 --> 00:00:55,333
k nearest neighbors classifier was created
and trained properly.

19
00:00:55,800 --> 00:00:57,133
And at the end.

20
00:00:57,133 --> 00:00:59,400
So all the predictions are generated.

21
00:00:59,400 --> 00:01:01,966
Again we can actually have a look at them.

22
00:01:01,966 --> 00:01:05,966
And wow, we seem to have some great
predictions that remember on the test set.

23
00:01:06,233 --> 00:01:07,633
So all this is correct.

24
00:01:07,633 --> 00:01:11,466
Here we have an incorrect prediction,
meaning that the K and then predicted

25
00:01:11,466 --> 00:01:12,733
this particular customer

26
00:01:12,733 --> 00:01:16,866
to buy the CV, whereas in reality
that customer didn't buy the SUV.

27
00:01:17,400 --> 00:01:19,600
Then all this is correct or correct?

28
00:01:19,600 --> 00:01:20,466
All correct.

29
00:01:20,466 --> 00:01:21,633
It seems really good.

30
00:01:21,633 --> 00:01:24,766
Here we have another incorrect prediction,
this time the opposite one.

31
00:01:24,933 --> 00:01:26,033
Our model predicted

32
00:01:26,033 --> 00:01:30,666
that this particular customer didn't buy
the SUV, whereas in reality that customer.

33
00:01:30,666 --> 00:01:32,066
But the SUV.

34
00:01:32,066 --> 00:01:34,500
All right. Correct. All correct.
All correct.

35
00:01:34,500 --> 00:01:35,233
So pretty good.

36
00:01:35,233 --> 00:01:38,400
Wow I think we actually have very,
very few incorrect prediction.

37
00:01:38,433 --> 00:01:40,700
Let's actually check it out right now
okay. No.

38
00:01:40,700 --> 00:01:42,266
So I didn't see it properly.

39
00:01:42,266 --> 00:01:47,066
But we actually have in total four
plus three incorrect predictions.

40
00:01:47,066 --> 00:01:48,133
That's really really good.

41
00:01:48,133 --> 00:01:50,300
That's better than the logistic regression
model.

42
00:01:50,300 --> 00:01:54,400
Remember we had actually
an accuracy of the test set

43
00:01:55,400 --> 00:01:57,566
of 89%.

44
00:01:57,566 --> 00:01:59,833
So here
clearly the K-nearest neighbors did

45
00:01:59,833 --> 00:02:01,500
better than the logistic regression model.

46
00:02:01,500 --> 00:02:05,466
And you will perfectly understand why
when I show you the visualization.

47
00:02:05,733 --> 00:02:07,333
So it's actually here.

48
00:02:07,333 --> 00:02:07,700
All right.

49
00:02:07,700 --> 00:02:10,933
So four plus three equals
seven incorrect predictions.

50
00:02:11,166 --> 00:02:14,666
Then 64 correct
predictions of the class zero

51
00:02:14,666 --> 00:02:18,966
meaning 64 correct predictions
that the customer doesn't buy the SUV.

52
00:02:19,266 --> 00:02:22,133
And 29 correct
predictions of the class one

53
00:02:22,133 --> 00:02:26,066
meaning 29 correct predictions
that the customer buys the SUV.

54
00:02:26,466 --> 00:02:27,433
And then we have indeed

55
00:02:27,433 --> 00:02:31,466
four incorrect predictions of the class
one, meaning four incorrect predictions.

56
00:02:31,533 --> 00:02:36,300
The customer buys the SUV and three
incorrect predictions of the class zero,

57
00:02:36,300 --> 00:02:40,866
meaning three incorrect predictions
that the customer doesn't buy the SUV.

58
00:02:41,366 --> 00:02:42,766
So very, very good.

59
00:02:42,766 --> 00:02:47,400
And now, as you can see,
it is still executing the cell that plots

60
00:02:47,700 --> 00:02:49,333
that 2D array.

61
00:02:49,333 --> 00:02:52,500
You know, showing
all the prediction regions

62
00:02:52,500 --> 00:02:55,700
and prediction
boundary of the canine model

63
00:02:55,800 --> 00:02:59,000
first on the training set
and then on the test set.

64
00:02:59,500 --> 00:03:03,133
And the reason why
it is taking some time here to execute.

65
00:03:03,133 --> 00:03:03,700
Oh, there we go.

66
00:03:03,700 --> 00:03:07,500
We have perfect timing is simply because,
you know, the K and end

67
00:03:07,500 --> 00:03:10,933
model is actually very compute intensive.

68
00:03:10,966 --> 00:03:14,466
You know, there are a lot of computations
when running the canine model.

69
00:03:14,466 --> 00:03:16,133
And that's why it took some time.

70
00:03:16,133 --> 00:03:18,466
And so now let's interpret the results.

71
00:03:18,466 --> 00:03:21,500
So remember with the logistic regression

72
00:03:21,500 --> 00:03:25,133
classifier the prediction
boundary was a straight line.

73
00:03:25,133 --> 00:03:29,100
And that's because the logistic regression
model is a linear classifier.

74
00:03:29,166 --> 00:03:32,166
We'll actually see another type
of linear classifier in this part.

75
00:03:32,333 --> 00:03:36,500
But remember for linear classifiers
the prediction boundary

76
00:03:36,500 --> 00:03:40,400
or prediction curve for classification
is a straight line.

77
00:03:40,400 --> 00:03:42,700
And in three dimensions
it's a straight plan.

78
00:03:42,700 --> 00:03:45,600
And then in dimension
it's a straight hyperplane.

79
00:03:45,600 --> 00:03:48,600
And so here we had a straight line
which therefore

80
00:03:48,633 --> 00:03:51,733
resulted
in having a lot of incorrect predictions.

81
00:03:51,733 --> 00:03:54,433
Because as we can see
there are a lot of green points here.

82
00:03:54,433 --> 00:03:58,800
You know, green customers
who but the SUV in reality that fell

83
00:03:58,800 --> 00:04:02,766
in the red region where we predict indeed
that the customers don't buy the SUV.

84
00:04:03,000 --> 00:04:04,300
And same for these ones here.

85
00:04:04,300 --> 00:04:08,333
You know, here we have a lot of customers
who in reality but the SUV.

86
00:04:08,333 --> 00:04:11,566
But we're predicting the two
because they fall in the red region.

87
00:04:11,933 --> 00:04:15,733
And so what we were hoping for
is to build a new classifier

88
00:04:15,966 --> 00:04:20,200
for which the decision boundary, you know,
separating the two prediction regions

89
00:04:20,433 --> 00:04:23,833
does something like this,
you know, would not be a straight line,

90
00:04:23,833 --> 00:04:27,133
but some kind of a curve catching
all the red points here

91
00:04:27,133 --> 00:04:30,100
without catching those green points
here and here.

92
00:04:30,100 --> 00:04:32,000
Okay, so that's what we were hoping for.

93
00:04:32,000 --> 00:04:34,266
And well, that's exactly what we got here.

94
00:04:34,266 --> 00:04:36,266
Thanks to our Kiernan model.

95
00:04:36,266 --> 00:04:40,300
We don't get a smooth curve,
but we get actually some kind of a curve

96
00:04:40,300 --> 00:04:42,766
that indeed catches
all the red points here

97
00:04:42,766 --> 00:04:46,466
without catching the green points
and leaving those green points

98
00:04:46,466 --> 00:04:50,600
and these ones inside the green region
where they should end up right.

99
00:04:50,966 --> 00:04:54,333
And of course, here we have some
green points that were super hard to catch

100
00:04:54,333 --> 00:04:57,166
because they are stuck in the middle
of all those red points here.

101
00:04:57,166 --> 00:04:57,766
So that's fine.

102
00:04:57,766 --> 00:05:01,433
And anyway, in machine learning
we want to avoid overfitting.

103
00:05:01,433 --> 00:05:04,466
You know, when we have some two perfect
predictions on the training set

104
00:05:04,600 --> 00:05:08,166
and therefore probably resulting
in having bad performance, bad predictions

105
00:05:08,366 --> 00:05:10,900
on the test set and well,
speaking of the test set,

106
00:05:10,900 --> 00:05:14,466
let's see how our key and did
it should be fine now.

107
00:05:14,500 --> 00:05:19,833
Yes, it is executed
and once again we get excellent results.

108
00:05:19,933 --> 00:05:20,233
Right.

109
00:05:20,233 --> 00:05:23,966
Because same the key and prediction
curve or prediction boundary

110
00:05:23,966 --> 00:05:28,100
or even decision boundary catches
all the right customers.

111
00:05:28,100 --> 00:05:30,366
You know, the red customers
who didn't buy in reality

112
00:05:30,366 --> 00:05:34,233
the SUV therefore
leaving the green customers who

113
00:05:34,233 --> 00:05:38,933
but in reality the SUV in the right
prediction region, the green region.

114
00:05:39,433 --> 00:05:41,833
All right. So that's already much better.

115
00:05:41,833 --> 00:05:46,233
And indeed we increased
the accuracy of the test set by 4%.

116
00:05:46,533 --> 00:05:49,800
And now the question is
can we do even better than this.

117
00:05:49,800 --> 00:05:52,233
Well, that's exactly what we're going
to find out

118
00:05:52,233 --> 00:05:55,666
in the next practical activities
to discover the big winner.

119
00:05:55,966 --> 00:05:57,833
And until then enjoy machine learning.