1
00:00:00,066 --> 00:00:03,033
And now
we're actually going to understand why.

2
00:00:03,033 --> 00:00:07,300
Why didn't the SVM model beat the K
and then.

3
00:00:07,566 --> 00:00:10,200
Okay. So we'll actually figure
that out in a bit.

4
00:00:10,200 --> 00:00:11,166
Or actually, you know,

5
00:00:11,166 --> 00:00:15,433
I can just show it to you here
on the original implementation of the SVM.

6
00:00:15,666 --> 00:00:18,900
But you're going to understand
now why it didn't beat it.

7
00:00:19,100 --> 00:00:20,333
Well, there you go.

8
00:00:20,333 --> 00:00:24,400
It didn't beat it because once again,
since we chose a linear kernel,

9
00:00:24,566 --> 00:00:26,366
well, the prediction boundary

10
00:00:26,366 --> 00:00:30,166
or you know, the decision boundary
is actually a once again a straight line.

11
00:00:30,166 --> 00:00:34,666
And therefore even if you rotate it either
this way, with this way, well,

12
00:00:34,666 --> 00:00:37,800
it won't be able to catch the right
predictions for, you know,

13
00:00:37,800 --> 00:00:41,100
these green customers here,
which should belong to the green regions.

14
00:00:41,100 --> 00:00:41,966
And same for this one.

15
00:00:41,966 --> 00:00:45,300
You know, if we rotated this way,
well we will catch these ones.

16
00:00:45,300 --> 00:00:48,300
But then we will catch
more incorrect predictions around here.

17
00:00:48,400 --> 00:00:51,900
And if we rotate this way,
you know, from here to here, well, indeed

18
00:00:51,900 --> 00:00:54,233
we will catch these ones in the right
prediction region,

19
00:00:54,233 --> 00:00:56,800
but we will get more incorrect ones here.

20
00:00:56,800 --> 00:00:59,733
So that's the problem of linear models
here.

21
00:00:59,733 --> 00:01:00,833
Indeed, it's much better

22
00:01:00,833 --> 00:01:04,800
to have a prediction boundary
that does some kind of a curve here

23
00:01:04,900 --> 00:01:09,066
in order to catch only the red customers
who should be predicted

24
00:01:09,100 --> 00:01:13,100
not by the SUV, and leave
the green ones in the green region.

25
00:01:13,200 --> 00:01:13,866
All right.

26
00:01:13,866 --> 00:01:17,900
That's exactly what we got
with our k nearest neighbors, right?

27
00:01:18,200 --> 00:01:22,133
It is not a smooth curve,
but it did the job of selecting the right

28
00:01:22,200 --> 00:01:23,700
red customers here.

29
00:01:23,700 --> 00:01:24,966
Still leaving some green ones,

30
00:01:24,966 --> 00:01:28,033
but here catching the right
green ones in the right region.

31
00:01:28,033 --> 00:01:28,966
The green region.

32
00:01:28,966 --> 00:01:32,166
And here of course, since we have this
straight line, it's impossible to do.

33
00:01:32,466 --> 00:01:34,000
But no worries.

34
00:01:34,000 --> 00:01:38,433
I'm sure you have the intuition
that once we use a nonlinear kernel,

35
00:01:38,466 --> 00:01:42,600
well, we'll be able to catch the right
green observations here in the right

36
00:01:42,600 --> 00:01:43,466
green region.

37
00:01:43,466 --> 00:01:47,300
And well, that's exactly what
we'll find out in the next section.

38
00:01:47,433 --> 00:01:50,433
When implementing the kernel SVM model.

39
00:01:50,900 --> 00:01:52,866
All right.
So let's see if we're done here.

40
00:01:52,866 --> 00:01:53,566
Yes we are.

41
00:01:53,566 --> 00:01:55,800
So it's finished to execute.

42
00:01:55,800 --> 00:01:57,566
And of course we get the same results.

43
00:01:57,566 --> 00:02:00,366
And let's
see for the test set. Still running. But

44
00:02:01,500 --> 00:02:03,600
and oh well okay.

45
00:02:03,600 --> 00:02:05,066
So funny timing.

46
00:02:05,066 --> 00:02:06,866
It just populated here

47
00:02:06,866 --> 00:02:09,866
and well you know it's still the same
because of this straight line here.

48
00:02:09,866 --> 00:02:12,600
Well we have some green customers
you know customers who.

49
00:02:12,600 --> 00:02:17,166
But in reality the new SUV but couldn't
be predicted correctly because they fall

50
00:02:17,166 --> 00:02:22,000
in the wrong region because of this
prediction boundary that cannot separate.

51
00:02:22,000 --> 00:02:24,466
Well are two classes all right.

52
00:02:24,466 --> 00:02:25,333
So same problem.

53
00:02:25,333 --> 00:02:29,000
And you know you might guess
that this problem will be fixed once

54
00:02:29,000 --> 00:02:32,700
we choose a nonlinear kernel
for our support vector machine model.

55
00:02:32,966 --> 00:02:33,300
All right.

56
00:02:33,300 --> 00:02:37,700
So let's find out about this
in the next section on kernel SVM.

57
00:02:37,733 --> 00:02:40,733
And until then enjoy machine learning.