/
test_moments_equalized_odds.py
201 lines (158 loc) · 7.13 KB
/
test_moments_equalized_odds.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
# Copyright (c) Microsoft Corporation. All rights reserved.
# Licensed under the MIT License.
import numpy as np
import pandas as pd
from fairlearn.reductions import EqualizedOdds
from fairlearn.reductions._moments.moment import _EVENT, _GROUP_ID, _SIGN
from .data_generator import simple_binary_threshold_data
def test_construct_and_load():
eqo = EqualizedOdds()
assert eqo.short_name == "EqualizedOdds"
num_samples_a0 = 10
num_samples_a1 = 30
num_samples = num_samples_a0 + num_samples_a1
a0_threshold = 0.2
a1_threshold = 0.7
a0_label = "a0"
a1_label = "a1"
X, Y, A = simple_binary_threshold_data(num_samples_a0, num_samples_a1,
a0_threshold, a1_threshold,
a0_label, a1_label)
# Load up the (rigged) data
eqo.load_data(X, Y, sensitive_features=A)
assert eqo.data_loaded
assert eqo.n == num_samples_a0 + num_samples_a1
# Examine the tags DF
assert eqo.tags['label'].equals(pd.Series(Y))
assert eqo.tags['group_id'].equals(pd.Series(A))
expected_tags_event = ['label={0}'.format(a) for a in Y]
assert np.array_equal(expected_tags_event, eqo.tags['event'])
# Examine the index MultiIndex
events = ['label=False', 'label=True']
signs = ['+', '-']
labels = [a0_label, a1_label]
expected_index = pd.MultiIndex.from_product(
[signs, events, labels],
names=[_SIGN, _EVENT, _GROUP_ID])
assert eqo.index.equals(expected_index)
# Examine the prob_event DF
# There are two events - 'True' and 'False'
assert len(eqo.prob_event.index) == 2
assert eqo.prob_event.loc['label=False'] == 1 - sum(Y) / len(Y)
assert eqo.prob_event.loc['label=True'] == sum(Y) / len(Y)
# Examine the prob_group_event DF
# There's only an 'all' event but this records the fractions
# of each label in the population
assert len(eqo.prob_group_event.index) == 4
# Use the fact that our data are uniformly distributed in the range [0, 1]
# With the current values, it appears we don't need to fiddle for off-by-one cases
a0_below = a0_threshold * num_samples_a0
a0_above = num_samples_a0 - a0_below
assert eqo.prob_group_event.loc[('label=False', a0_label)] == a0_below / num_samples
assert eqo.prob_group_event.loc[('label=True', a0_label)] == a0_above / num_samples
a1_below = a1_threshold * num_samples_a1
a1_above = num_samples_a1 - a1_below
assert eqo.prob_group_event.loc[('label=False', a1_label)] == a1_below / num_samples
assert eqo.prob_group_event.loc[('label=True', a1_label)] == a1_above / num_samples
# Examine the neg_basis DF
assert len(eqo.neg_basis.index) == 8
assert eqo.neg_basis[0]['+', 'label=False', a0_label] == 0
assert eqo.neg_basis[0]['+', 'label=False', a1_label] == 0
assert eqo.neg_basis[0]['+', 'label=True', a0_label] == 0
assert eqo.neg_basis[0]['+', 'label=True', a1_label] == 0
assert eqo.neg_basis[0]['-', 'label=False', a0_label] == 1
assert eqo.neg_basis[0]['-', 'label=False', a1_label] == 0
assert eqo.neg_basis[0]['-', 'label=True', a0_label] == 0
assert eqo.neg_basis[0]['-', 'label=True', a1_label] == 0
# Examine the pos_basis DF
# This is looking at the \lambda_{+} values and picking out the
# one associated with the first label
assert len(eqo.pos_basis.index) == 8
assert eqo.pos_basis[0]['+', 'label=False', a0_label] == 1
assert eqo.pos_basis[0]['+', 'label=False', a1_label] == 0
assert eqo.pos_basis[0]['+', 'label=True', a0_label] == 0
assert eqo.pos_basis[0]['+', 'label=True', a1_label] == 0
assert eqo.pos_basis[0]['-', 'label=False', a0_label] == 0
assert eqo.pos_basis[0]['-', 'label=False', a1_label] == 0
assert eqo.pos_basis[0]['-', 'label=True', a0_label] == 0
assert eqo.pos_basis[0]['-', 'label=True', a1_label] == 0
# Examine the neg_basis_present DF
assert eqo.neg_basis_present[0]
def test_project_lambda_smoke_negatives():
eqo = EqualizedOdds()
events = ['label=False', 'label=True']
signs = ['+', '-']
labels = ['a', 'b']
midx = pd.MultiIndex.from_product(
[signs, events, labels],
names=[_SIGN, _EVENT, _GROUP_ID])
df = pd.DataFrame()
# Note that the '-' labels are larger
df = 0 + pd.Series([1, 2, 11, 19, 1001, 1110, 1230, 1350], index=midx)
ls = eqo.project_lambda(df)
expected = pd.DataFrame()
expected = 0 + pd.Series([0, 0, 0, 0, 1000, 1108, 1219, 1331], index=midx)
assert expected.equals(ls)
def test_project_lambda_smoke_positives():
# This is a repeat of the _negatives method but with
# the '+' indices larger
eqo = EqualizedOdds()
events = ['label=False', 'label=True']
signs = ['+', '-']
labels = ['a', 'b']
midx = pd.MultiIndex.from_product(
[signs, events, labels],
names=[_SIGN, _EVENT, _GROUP_ID])
df = pd.DataFrame()
# Note that the '-' indices are now smaller
df = 0 + pd.Series([200, 300, 100, 600, 4, 5, 6, 7], index=midx)
ls = eqo.project_lambda(df)
expected = pd.DataFrame()
expected = 0 + pd.Series([196, 295, 94, 593, 0, 0, 0, 0], index=midx)
assert expected.equals(ls)
def test_signed_weights():
eqo = EqualizedOdds()
assert eqo.short_name == "EqualizedOdds"
num_samples_a0 = 10
num_samples_a1 = 30
num_samples = num_samples_a0 + num_samples_a1
a0_threshold = 0.2
a1_threshold = 0.7
a0_label = "OneThing"
a1_label = "AnotherThing"
X, Y, A = simple_binary_threshold_data(num_samples_a0, num_samples_a1,
a0_threshold, a1_threshold,
a0_label, a1_label)
# Load up the (rigged) data
eqo.load_data(X, Y, sensitive_features=A)
events = ['label=False', 'label=True']
signs = ["+", "-"]
labels = [a0_label, a1_label]
midx = pd.MultiIndex.from_product(
[signs, events, labels],
names=[_SIGN, _EVENT, _GROUP_ID])
lambda_vec = pd.Series([2000, 1000, 4000, 5000, 500, 100, 700, 900], index=midx, name=0)
lambda_a0_F = 2000 - 500
lambda_a0_T = 4000 - 700
num_a0_F = int(a0_threshold * num_samples_a0)
num_a0_T = num_samples_a0 - num_a0_F
lambda_a1_F = 1000 - 100
lambda_a1_T = 5000 - 900
num_a1_F = int(a1_threshold * num_samples_a1)
num_a1_T = num_samples_a1 - num_a1_F
sw_a0_F = (lambda_a0_F + lambda_a1_F) / (1 - sum(Y) / len(Y)) - \
lambda_a0_F * (num_samples / num_a0_F)
sw_a1_F = (lambda_a0_F + lambda_a1_F) / (1 - sum(Y) / len(Y)) - \
lambda_a1_F * (num_samples / num_a1_F)
sw_a0_T = (lambda_a0_T + lambda_a1_T) / (sum(Y) / len(Y)) - \
lambda_a0_T * (num_samples / num_a0_T)
sw_a1_T = (lambda_a0_T + lambda_a1_T) / (sum(Y) / len(Y)) - \
lambda_a1_T * (num_samples / num_a1_T)
w_a0_F = np.full(num_a0_F, sw_a0_F)
w_a0_T = np.full(num_a0_T, sw_a0_T)
w_a1_F = np.full(num_a1_F, sw_a1_F)
w_a1_T = np.full(num_a1_T, sw_a1_T)
expected = np.concatenate((w_a0_F, w_a0_T, w_a1_F, w_a1_T), axis=None)
signed_weights = eqo.signed_weights(lambda_vec)
# Be bold and test for equality
assert np.array_equal(expected, signed_weights)