1
00:00:00,042 --> 00:00:04,689
[MUSIC]

2
00:00:04,689 --> 00:00:08,275
In my opinion the best thing you can
to do to deal with missing data is

3
00:00:08,275 --> 00:00:12,644
to make your algorithm, machine learning
algorithm, robust to missing data.

4
00:00:13,715 --> 00:00:16,315
In other words, make sure that
an algorithm adapt to missing data.

5
00:00:16,315 --> 00:00:19,745
And that's exactly what we're going to
do in this particular module.

6
00:00:19,745 --> 00:00:23,195
And we're going to do it in
the context of decision trees,

7
00:00:23,195 --> 00:00:26,685
because a simple modification of decision
trees, they can handle missing data.

8
00:00:27,685 --> 00:00:30,445
So, how we going to
deal with missing data?

9
00:00:30,445 --> 00:00:33,180
Let us try to understand a little better

10
00:00:33,180 --> 00:00:35,330
what happened in the context
of the decision trees?

11
00:00:35,330 --> 00:00:40,840
So I have this input xi where the credit
was poor and income was unknown.

12
00:00:40,840 --> 00:00:45,920
I go down the decision tree,
I hit credit poor, so

13
00:00:45,920 --> 00:00:51,270
I take the branch here, poor.

14
00:00:51,270 --> 00:00:54,075
And then I hit income,
but income was unknown.

15
00:00:54,075 --> 00:00:56,250
Income was question mark.

16
00:00:56,250 --> 00:00:57,050
So what do I do then?

17
00:00:57,050 --> 00:01:02,030
So what we're going to do is
assign a branch to follow

18
00:01:02,030 --> 00:01:05,180
when the data is unknown,
when the value is unknown.

19
00:01:05,180 --> 00:01:08,870
And in this case, I'm going to
associate all unknown data, for

20
00:01:08,870 --> 00:01:13,130
example, with the branch
where income was low.

21
00:01:13,130 --> 00:01:19,130
So if your income was unknown,
take the branch low and

22
00:01:19,130 --> 00:01:22,890
from there I'm just going to
predict that this loan was risky.

23
00:01:22,890 --> 00:01:28,280
So in other words, in our decision tree
we're going to make explicit decisions

24
00:01:28,280 --> 00:01:32,770
as to where question marks or
missing values, unknown values will go.

25
00:01:34,560 --> 00:01:38,930
We've introduced a decision to what
happens when we hit an unknown value

26
00:01:38,930 --> 00:01:40,810
at this point of the decision tree.

27
00:01:40,810 --> 00:01:43,940
But we may have unknown values any point
of decision tree, so we're going to

28
00:01:43,940 --> 00:01:48,990
make decisions everywhere as to what
should happen when we see unknown values.

29
00:01:48,990 --> 00:01:53,960
So for every decision known, when I
choose one of them its right to absorb

30
00:01:54,980 --> 00:01:59,730
the unknown values and
notice that those are going to be specific

31
00:01:59,730 --> 00:02:03,060
decisions we're going to make
associated with every diamond known or

32
00:02:03,060 --> 00:02:08,240
every decision known in the decision
tree so for credit we're going to have

33
00:02:08,240 --> 00:02:13,335
the unknowns or question marks
going to the fair branch for

34
00:02:13,335 --> 00:02:17,600
income, it was poor.

35
00:02:17,600 --> 00:02:23,300
And we look at income,
we're going to go down the lower branch.

36
00:02:23,300 --> 00:02:27,980
When credit equals poor and
income is high we look at term and,

37
00:02:27,980 --> 00:02:31,830
in this case,
when term is equal to five years,

38
00:02:31,830 --> 00:02:37,430
we're going to put the unknowns, or the
question marks, in the five year bunch.

39
00:02:37,430 --> 00:02:40,720
However, note take a note of this,

40
00:02:40,720 --> 00:02:43,530
we might choose to say that
when credit goes fair and

41
00:02:43,530 --> 00:02:49,860
we look at term the unknowns in this case
will go to the term equals three years.

42
00:02:49,860 --> 00:02:55,080
So the decision that we make
about where unknowns go

43
00:02:55,080 --> 00:02:58,720
does not need to be the same in
different parts of the decision tree.

44
00:02:59,820 --> 00:03:01,871
So we take different decisions here,

45
00:03:04,529 --> 00:03:09,130
About the questions marks, or unknowns.

46
00:03:10,300 --> 00:03:13,043
And that's the beauty of
the approach that we're describing,

47
00:03:13,043 --> 00:03:17,158
because we're going to optimize, just like
we're optimizing of the decision trees,

48
00:03:17,158 --> 00:03:19,443
we're going to optimize
where the unknowns will go,

49
00:03:19,443 --> 00:03:21,558
in order to minimize
the classification error.

50
00:03:21,558 --> 00:03:27,690
Now if a learning tree like this and
we're given an input,

51
00:03:27,690 --> 00:03:30,930
for example, where the credit is
unknown and the income was high,

52
00:03:30,930 --> 00:03:34,700
and the term was 5 years,
we can go down this tree and traverse.

53
00:03:34,700 --> 00:03:38,060
Well, credit was unknown,
the term was 5 years, so

54
00:03:38,060 --> 00:03:39,940
I'm going to predict it's a safe loan.

55
00:03:42,180 --> 00:03:45,720
While if I have another input
where the credit is poor,

56
00:03:45,720 --> 00:03:50,980
the income is high, the term was
question mark when the different

57
00:03:50,980 --> 00:03:55,510
branch of the decision tree will go
down credit poor, income high and

58
00:03:57,190 --> 00:04:02,410
the term here unknown goes down
the same as the five year loan and

59
00:04:02,410 --> 00:04:05,680
again we'll predict safe.

60
00:04:05,680 --> 00:04:09,990
In general, approaches that
explicitly handle missing data

61
00:04:09,990 --> 00:04:13,010
in the algorithm itself
can be quite helpful.

62
00:04:13,010 --> 00:04:18,150
They can help address missing data
training time and prediction time.

63
00:04:18,150 --> 00:04:20,990
And can make more accurate
predictions in general.

64
00:04:20,990 --> 00:04:23,760
And we just talked about
an example in decision trees.

65
00:04:23,760 --> 00:04:27,930
The downside though is that it requires it
to modify the actual learning algorithm to

66
00:04:27,930 --> 00:04:29,640
deal with missing data.

67
00:04:29,640 --> 00:04:32,190
In the case of decision trees,
I'm going to describe

68
00:04:32,190 --> 00:04:35,270
kind of a very simple modification
that can make this happen,

69
00:04:35,270 --> 00:04:37,580
but this can be very complex for
other algorithms.

70
00:04:37,580 --> 00:04:39,760
And even for decision trees, if you're
going to make it really, really good,

71
00:04:39,760 --> 00:04:42,530
it might be more complex than
what we're talking about next.

72
00:04:42,530 --> 00:04:44,830
But the idea here is
fundamentally important.

73
00:04:44,830 --> 00:04:45,330
[MUSIC]