1
00:00:00,000 --> 00:00:03,952
[MUSIC PLAYING]

2
00:00:05,762 --> 00:00:07,220
MIKE TEODORESCU:
Hello, and welcome

3
00:00:07,220 --> 00:00:09,280
to this module on choices
of fairness criteria.

4
00:00:09,280 --> 00:00:10,210
My name is Mike Teodorescu.

5
00:00:10,210 --> 00:00:11,860
I'm an assistant professor
of information systems

6
00:00:11,860 --> 00:00:14,277
at Boston College, as was a
visiting scholar at MIT D-Lab.

7
00:00:14,277 --> 00:00:15,777
What we'll cover
in this module will

8
00:00:15,777 --> 00:00:17,290
be the concept of
confusion matrix,

9
00:00:17,290 --> 00:00:20,470
as well as three popular
examples of fairness criteria--

10
00:00:20,470 --> 00:00:22,600
demographic parity,
equalized odds,

11
00:00:22,600 --> 00:00:25,418
and equalized opportunity.

12
00:00:25,418 --> 00:00:26,960
Some countries have
laws that protect

13
00:00:26,960 --> 00:00:29,270
specific groups of people
from discrimination

14
00:00:29,270 --> 00:00:31,490
based on certain attributes.

15
00:00:31,490 --> 00:00:33,480
As we review in
the previous video,

16
00:00:33,480 --> 00:00:35,630
these are called
protected attributes.

17
00:00:35,630 --> 00:00:38,523
Some examples are on this slide.

18
00:00:38,523 --> 00:00:39,940
Regardless of the
legal landscape,

19
00:00:39,940 --> 00:00:41,357
machine learning
has the potential

20
00:00:41,357 --> 00:00:45,020
to produce unfair outcomes
for certain groups of people.

21
00:00:45,020 --> 00:00:47,540
As an algorithm designer,
one should make clear choices

22
00:00:47,540 --> 00:00:49,520
about fairness criteria.

23
00:00:49,520 --> 00:00:51,980
Some criteria will be reviewed
in the next few slides.

24
00:00:54,285 --> 00:00:56,910
[? In the ?] previous video, we
also discussed case of fairness

25
00:00:56,910 --> 00:00:58,910
through unawareness,
which refers

26
00:00:58,910 --> 00:01:02,190
to leaving out the protected
attributes out of your model.

27
00:01:02,190 --> 00:01:04,448
And we also explained why
this is not a good choice.

28
00:01:04,448 --> 00:01:06,740
Specifically, you may end up
with other attributes that

29
00:01:06,740 --> 00:01:08,630
correlate with
protected attributes,

30
00:01:08,630 --> 00:01:11,300
and you may end up
discriminating inadvertently

31
00:01:11,300 --> 00:01:13,928
nonetheless.

32
00:01:13,928 --> 00:01:15,970
In order to go into
additional fairness criteria,

33
00:01:15,970 --> 00:01:18,750
we need to discuss some
additional concepts.

34
00:01:18,750 --> 00:01:21,750
Consider you have a
binary classifier.

35
00:01:21,750 --> 00:01:24,630
For example, you're
looking at decision of hire

36
00:01:24,630 --> 00:01:28,380
or not hired or to lend
credit and not to lend credit.

37
00:01:28,380 --> 00:01:31,957
If we want to look at the
predictions for a model that

38
00:01:31,957 --> 00:01:33,540
would do such a
binary classification,

39
00:01:33,540 --> 00:01:35,420
we would look at the
predicted values.

40
00:01:35,420 --> 00:01:38,292
We could bucket them in four
categories-- true positives,

41
00:01:38,292 --> 00:01:39,750
which would be
correctly classified

42
00:01:39,750 --> 00:01:42,300
as positive, true negative,
correctly classified

43
00:01:42,300 --> 00:01:45,645
as negative, false positives,
which would be values that

44
00:01:45,645 --> 00:01:48,338
were incorrectly classified
as positive by the algorithm,

45
00:01:48,338 --> 00:01:50,880
false negatives, which would be
values incorrectly classified

46
00:01:50,880 --> 00:01:52,240
as negative.

47
00:01:52,240 --> 00:01:54,285
And, if we were to add
the the true positives

48
00:01:54,285 --> 00:01:58,270
to the true negatives and
divide that by all four,

49
00:01:58,270 --> 00:02:01,530
we would end up with the value
of the accuracy of the model.

50
00:02:01,530 --> 00:02:04,650
In this example where
accuracy is this fraction,

51
00:02:04,650 --> 00:02:08,970
an accuracy of 0.5 is the same
as a random classification.

52
00:02:08,970 --> 00:02:11,980
Now we should look
at accuracy carefully

53
00:02:11,980 --> 00:02:14,760
and see that it doesn't tell us
anything about the prediction

54
00:02:14,760 --> 00:02:15,360
of negatives.

55
00:02:15,360 --> 00:02:17,790
It could mislead us if two
classes were imbalanced,

56
00:02:17,790 --> 00:02:21,060
for example, if 90% of
the sample is positives,

57
00:02:21,060 --> 00:02:22,620
and 10% is negatives.

58
00:02:22,620 --> 00:02:24,990
For that, we have other
additional criteria

59
00:02:24,990 --> 00:02:29,130
we could go into deeper, such
as MCC score and AUC score.

60
00:02:31,940 --> 00:02:34,130
The true positives, true
negatives, false positives,

61
00:02:34,130 --> 00:02:37,410
and false negatives are,
oftentimes, represented in a 2

62
00:02:37,410 --> 00:02:41,160
by 2 matrix form called
a confusion matrix.

63
00:02:41,160 --> 00:02:43,560
This is simply an easier
presentation for us

64
00:02:43,560 --> 00:02:47,870
to see the behavior
of the classifier.

65
00:02:47,870 --> 00:02:49,670
The first additional
fairness criteria

66
00:02:49,670 --> 00:02:51,900
is called demographic parity.

67
00:02:51,900 --> 00:02:55,560
It's a criterion for what's
called group-level fairness.

68
00:02:55,560 --> 00:02:58,260
This criterion specify
that the outcome, which

69
00:02:58,260 --> 00:03:00,660
here is denoted by a
y hat, is independent

70
00:03:00,660 --> 00:03:05,310
of the protected attribute A.
For example, the probability

71
00:03:05,310 --> 00:03:09,610
of being hired is
independent of the gender.

72
00:03:09,610 --> 00:03:12,130
There are multiple problems
with demographic parity.

73
00:03:12,130 --> 00:03:13,810
One would be the
definition that we just

74
00:03:13,810 --> 00:03:16,030
discussed would
not hold if we had

75
00:03:16,030 --> 00:03:19,300
individuals who would be members
of multiple protected groups.

76
00:03:19,300 --> 00:03:21,642
By enforcing group-level
fairness for one attribute,

77
00:03:21,642 --> 00:03:23,350
we would still violate
the group fairness

78
00:03:23,350 --> 00:03:25,750
for other attributes or
combinations of attributes,

79
00:03:25,750 --> 00:03:29,310
such as subgroups
of the population.

80
00:03:29,310 --> 00:03:31,560
Furthermore, while enforcing
group-level fairness,

81
00:03:31,560 --> 00:03:34,620
for example, the same hiring
grade for females and males,

82
00:03:34,620 --> 00:03:36,737
that could still be
unfair to individuals.

83
00:03:36,737 --> 00:03:39,070
It could force the algorithm
to drop otherwise qualified

84
00:03:39,070 --> 00:03:40,800
individuals just to
achieve independence

85
00:03:40,800 --> 00:03:44,700
of outcome of the attribute.

86
00:03:44,700 --> 00:03:46,770
Fairness at the group
level could, potentially,

87
00:03:46,770 --> 00:03:49,300
be unfair at the
individual level.

88
00:03:49,300 --> 00:03:52,800
For example, if we have a
high rate of false positives,

89
00:03:52,800 --> 00:03:54,990
we could end up-- and a low
rate of false negatives,

90
00:03:54,990 --> 00:03:57,120
it could still end up
being unfair to individuals

91
00:03:57,120 --> 00:04:00,420
in that we could hire
people who are without merit

92
00:04:00,420 --> 00:04:03,030
at the disadvantage
of other individuals

93
00:04:03,030 --> 00:04:05,790
who could be qualified
and should be hired,

94
00:04:05,790 --> 00:04:08,550
but, due to the group-level
fairness criterion,

95
00:04:08,550 --> 00:04:11,160
we have to hire some
who are not qualified

96
00:04:11,160 --> 00:04:13,820
from one of the groups.

97
00:04:13,820 --> 00:04:15,710
The sweet spot would
be low false negatives

98
00:04:15,710 --> 00:04:18,440
and low false positives, which
would be fair, potentially,

99
00:04:18,440 --> 00:04:21,019
to both the individual
and the group level.

100
00:04:21,019 --> 00:04:23,240
We could also end up in
the top right corner, which

101
00:04:23,240 --> 00:04:26,060
would be the worst-case
scenario, low accuracy,

102
00:04:26,060 --> 00:04:28,550
unfair to individuals, but
potentially fair for the group

103
00:04:28,550 --> 00:04:30,410
where we have high
false negatives

104
00:04:30,410 --> 00:04:34,210
and high false positives.

105
00:04:34,210 --> 00:04:36,460
A stricter criterion
is equalized odds,

106
00:04:36,460 --> 00:04:38,770
which means matching both
of true positive rate

107
00:04:38,770 --> 00:04:41,200
and the false positive
rate for different values

108
00:04:41,200 --> 00:04:43,150
of the protected attribute.

109
00:04:43,150 --> 00:04:45,760
This is much harder to achieve
than demographic parity,

110
00:04:45,760 --> 00:04:49,150
but it is one of the higher
levels of algorithmic fairness.

111
00:04:49,150 --> 00:04:50,860
In this case, rather
than predicting

112
00:04:50,860 --> 00:04:52,690
the same average
risk across subgroups

113
00:04:52,690 --> 00:04:56,460
of protected social attributes,
like in demographic parity,

114
00:04:56,460 --> 00:04:59,080
equalized odds is stricter
in that it forces equality

115
00:04:59,080 --> 00:05:02,260
only among individuals who
reach similar outcomes.

116
00:05:02,260 --> 00:05:04,360
In the hiring example
that we've discussed

117
00:05:04,360 --> 00:05:06,640
in the previous
video, this implies

118
00:05:06,640 --> 00:05:09,120
that the probability of a
qualified applicant to be hired

119
00:05:09,120 --> 00:05:11,140
and the probability of
an unqualified applicant

120
00:05:11,140 --> 00:05:13,930
to be incorrectly hired should
be the same across genders.

121
00:05:17,030 --> 00:05:18,740
A milder version
of equalized odds

122
00:05:18,740 --> 00:05:20,510
is equalized opportunity
where we're only

123
00:05:20,510 --> 00:05:23,450
concerned with treating fairly
those who are determined

124
00:05:23,450 --> 00:05:26,540
to be worthy of acceptance,
i.e. dependent variable

125
00:05:26,540 --> 00:05:29,030
is equal to 1, or
they're worthy of being

126
00:05:29,030 --> 00:05:32,350
hired, worthy of being
admitted, and so on.

127
00:05:32,350 --> 00:05:34,850
Equalized opportunity is not
concerned with rejecting people

128
00:05:34,850 --> 00:05:37,140
fairly across protected groups.

129
00:05:37,140 --> 00:05:39,660
So to speak, the false positive
rates and the true positive

130
00:05:39,660 --> 00:05:41,952
rates do not both need to be
equal across the protected

131
00:05:41,952 --> 00:05:42,830
categories.

132
00:05:42,830 --> 00:05:44,300
We only need the
true positive rate

133
00:05:44,300 --> 00:05:47,410
to be equal across
protected categories.

134
00:05:47,410 --> 00:05:49,060
In a way, equalized
opportunity is

135
00:05:49,060 --> 00:05:51,280
less interventionist
than equalized odds,

136
00:05:51,280 --> 00:05:53,710
and it may be more
achievable, depending

137
00:05:53,710 --> 00:05:58,720
on your individual situation
and implementation challenges.

138
00:05:58,720 --> 00:06:01,330
In the example of
hiring, we only

139
00:06:01,330 --> 00:06:04,630
are concerned with individuals
who are worthy of being hired,

140
00:06:04,630 --> 00:06:07,000
i.e. the actual
qualified applicants.

141
00:06:07,000 --> 00:06:08,530
And, out of the
rejected applicants,

142
00:06:08,530 --> 00:06:10,155
we may be sometimes
rejecting unfairly.

143
00:06:12,610 --> 00:06:15,950
Here's some review questions
for the last two videos.

144
00:06:15,950 --> 00:06:17,960
What is demographic parity?

145
00:06:17,960 --> 00:06:20,580
What is fairness
through unawareness?

146
00:06:20,580 --> 00:06:23,490
Is fairness at the group
level always the best?

147
00:06:23,490 --> 00:06:25,915
What is a confusion matrix?

148
00:06:25,915 --> 00:06:27,540
What is the equality
of odds criterion?

149
00:06:27,540 --> 00:06:28,957
And when might you
want to use it?

150
00:06:31,150 --> 00:06:35,520
This course is sponsored by the
USAID grant through MIT D-Lab.

151
00:06:35,520 --> 00:06:38,160
And this is joint work with my
faculty colleagues Lily Morse

152
00:06:38,160 --> 00:06:41,400
and Gerald Kane from Boston
college and research assistant

153
00:06:41,400 --> 00:06:42,670
Yazeed Awwad from MIT D-Lab.

154
00:06:45,158 --> 00:06:46,950
If you would like to
learn more about this,

155
00:06:46,950 --> 00:06:50,040
please consult the
following references.

156
00:06:50,040 --> 00:06:51,960
Thank you so much for
watching this video.

157
00:06:51,960 --> 00:06:53,877
We hope you find it
useful and you'll continue

158
00:06:53,877 --> 00:06:55,880
watching the rest of the class.

159
00:06:55,880 --> 00:06:59,830
[MUSIC PLAYING]