﻿1
00:00:02,100 --> 00:00:04,200
Here I've got a confusion matrix.

2
00:00:04,200 --> 00:00:07,233
And this confusion
matrix has 10,000 records in it.

3
00:00:07,600 --> 00:00:10,600
And it represents scenario number one
which we'll be looking at.

4
00:00:11,100 --> 00:00:15,400
As you can see this model has made 150
type one errors and 50 type two errors.

5
00:00:15,733 --> 00:00:18,733
And but overall is predicted
quite a lot correctly.

6
00:00:18,933 --> 00:00:21,933
Now let's calculate
the accuracy rate in this scenario.

7
00:00:22,200 --> 00:00:24,566
The accuracy rate is the total correct
divide.

8
00:00:24,566 --> 00:00:29,466
But overall total and it's 9800
divided by 10,000 which is 98%.

9
00:00:30,133 --> 00:00:30,766
Okay great.

10
00:00:30,766 --> 00:00:32,300
But what are we going to do now?

11
00:00:32,300 --> 00:00:35,866
Is we going to tell the model
to stop making predictions

12
00:00:35,866 --> 00:00:38,866
whatsoever, which is going to abandon
the model completely?

13
00:00:38,933 --> 00:00:42,433
And we're going to say that from now on,
our prediction is always zero.

14
00:00:42,433 --> 00:00:46,000
We're always going to predict
that the event is not going to occur.

15
00:00:46,333 --> 00:00:49,200
So basically what will happen
to the confusion matrix is

16
00:00:49,200 --> 00:00:52,200
these records will move
from the right column to the left column.

17
00:00:52,366 --> 00:00:57,900
And our new confusion
matrix will look like this 9850 150.

18
00:00:57,900 --> 00:01:01,766
And then nothing in the predicted column
where we predicted

19
00:01:01,766 --> 00:01:03,600
that something will occur.

20
00:01:03,600 --> 00:01:05,800
And of course, that move
is against all logic, right?

21
00:01:05,800 --> 00:01:07,666
Why would you abandon a model?

22
00:01:07,666 --> 00:01:10,666
But let's calculate
the accuracy rate in this scenario.

23
00:01:10,800 --> 00:01:13,200
Scenario
two accuracy rate has the same formula.

24
00:01:13,200 --> 00:01:17,800
In this case, accuracy rate is 9850
divided by 10,000.

25
00:01:18,200 --> 00:01:20,800
So it's 98.5%.

26
00:01:20,800 --> 00:01:23,800
The accuracy rate
went up by half a percent.

27
00:01:24,300 --> 00:01:29,200
And as you can see, what we did is
we just completely stopped using a model.

28
00:01:29,200 --> 00:01:31,466
But the accuracy rate went up and

29
00:01:32,400 --> 00:01:34,633
that is why you should not

30
00:01:34,633 --> 00:01:39,933
base your judgment just on accuracy rate,
because things like this can happen.

31
00:01:39,933 --> 00:01:44,200
And even though obviously
you're not using a model anymore,

32
00:01:44,200 --> 00:01:48,266
which means that,
you're not applying any kind of logic

33
00:01:48,600 --> 00:01:52,633
into your decision making process,
your accuracy rate is going up.

34
00:01:52,633 --> 00:01:55,800
So it's misleading you in, into a wrong

35
00:01:55,800 --> 00:01:58,800
conclusion
that you should stop using models.

36
00:01:58,833 --> 00:02:01,566
And this effect is called
the accuracy paradox.

37
00:02:01,566 --> 00:02:05,100
And starting from the next tutorial,
I will show you a much better way

38
00:02:05,100 --> 00:02:09,100
to assess your models
using the cumulative accuracy profile.

39
00:02:09,633 --> 00:02:10,566
I look forward to seeing you then.

40
00:02:10,566 --> 00:02:13,566
Until next time, happy analyzing!
