Upload bias_auc.py
Browse filesAdded bias auc metric
- bias_auc.py +135 -0
bias_auc.py
ADDED
@@ -0,0 +1,135 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import evaluate
|
2 |
+
import datasets
|
3 |
+
from datasets.features import Sequence, Value, ClassLabel
|
4 |
+
from sklearn.metrics import roc_auc_score
|
5 |
+
import numpy as np
|
6 |
+
|
7 |
+
|
8 |
+
_DESCRIPTION = """\
|
9 |
+
Suite of threshold-agnostic metrics that provide a nuanced view
|
10 |
+
of this unintended bias, by considering the various ways that a
|
11 |
+
classifier’s score distribution can vary across designated groups.
|
12 |
+
|
13 |
+
The following are computed:
|
14 |
+
|
15 |
+
- BNSP (Background Negative, Subgroup Positive); and
|
16 |
+
- BPSN (Background Positive, Subgroup Negative) AUC metrics
|
17 |
+
|
18 |
+
"""
|
19 |
+
|
20 |
+
_CITATION = """\
|
21 |
+
@inproceedings{borkan2019nuanced,
|
22 |
+
title={Nuanced metrics for measuring unintended bias with real data for text classification},
|
23 |
+
author={Borkan, Daniel and Dixon, Lucas and Sorensen, Jeffrey and Thain, Nithum and Vasserman, Lucy},
|
24 |
+
booktitle={Companion proceedings of the 2019 world wide web conference},
|
25 |
+
pages={491--500},
|
26 |
+
year={2019}
|
27 |
+
}
|
28 |
+
"""
|
29 |
+
|
30 |
+
_KWARGS_DESCRIPTION = """\
|
31 |
+
target list[list[str]]: list containing list of group targeted for each item
|
32 |
+
label list[int]: list containing label index for each item
|
33 |
+
output list[list[float]]: list of model output values for each
|
34 |
+
"""
|
35 |
+
|
36 |
+
class BiasAUC(evaluate.EvaluationModule):
|
37 |
+
def _info(self):
|
38 |
+
return datasets.MetricInfo(
|
39 |
+
description=_DESCRIPTION,
|
40 |
+
citation=_CITATION,
|
41 |
+
inputs_description=_KWARGS_DESCRIPTION,
|
42 |
+
features=datasets.Features(
|
43 |
+
{
|
44 |
+
'target': Sequence(feature=Value(dtype='string', id=None), length=-1, id=None),
|
45 |
+
'label': Value(dtype='int64', id=None),
|
46 |
+
'output': Sequence(feature=Value(dtype='float32', id=None), length=-1, id=None),
|
47 |
+
}
|
48 |
+
),
|
49 |
+
reference_urls=["https://arxiv.org/abs/1903.04561"],
|
50 |
+
)
|
51 |
+
|
52 |
+
def _genreate_subgroup(self, targets, labels, outputs, subgroup, target_class=None):
|
53 |
+
"""Returns label and output score from `targets` and `labels`
|
54 |
+
if `subgroup` is in list of targeted groups found in `targets`
|
55 |
+
"""
|
56 |
+
target_class = target_class if target_class is not None else np.asarray(outputs).shape[-1] - 1
|
57 |
+
for target, label, result in zip(targets, labels, outputs):
|
58 |
+
if subgroup in target:
|
59 |
+
yield label, result[target_class]
|
60 |
+
|
61 |
+
def _genreate_bpsn(self, targets, labels, outputs, subgroup, target_class=None):
|
62 |
+
"""Returns label and output score from `targets` and `labels`
|
63 |
+
if (1) `subgroup` is in list of targeted groups found in `targets` and
|
64 |
+
label is not the same as `target_class`; or (2) `subgroup` is not in list of
|
65 |
+
targeted groups found in `targets` and label is the same as `target_class`
|
66 |
+
"""
|
67 |
+
target_class = target_class if target_class is not None else np.asarray(outputs).shape[-1] - 1
|
68 |
+
for target, label, result in zip(targets, labels, outputs):
|
69 |
+
if not target:
|
70 |
+
continue
|
71 |
+
# background positive
|
72 |
+
if subgroup not in target and label == target_class:
|
73 |
+
yield label, result[target_class]
|
74 |
+
# subgroup negative
|
75 |
+
elif subgroup in target and label != target_class:
|
76 |
+
yield label, result[target_class]
|
77 |
+
|
78 |
+
def _genreate_bnsp(self, targets, labels, outputs, subgroup, target_class=None):
|
79 |
+
"""Returns label and output score from `targets` and `labels`
|
80 |
+
if (1) `subgroup` is not in list of targeted groups found in `targets` and
|
81 |
+
label is the same as `target_class`; or (2) `subgroup` is in list of
|
82 |
+
targeted groups found in `targets` and label is not the same as `target_class`
|
83 |
+
"""
|
84 |
+
# get the index from class
|
85 |
+
target_class = target_class if target_class is not None else np.asarray(outputs).shape[-1] - 1
|
86 |
+
for target, label, result in zip(targets, labels, outputs):
|
87 |
+
if not target:
|
88 |
+
continue
|
89 |
+
# background negative
|
90 |
+
if subgroup not in target and label != target_class:
|
91 |
+
yield label, result[target_class]
|
92 |
+
# subgroup positive
|
93 |
+
elif subgroup in target and label == target_class:
|
94 |
+
yield label, result[target_class]
|
95 |
+
|
96 |
+
def _auc_by_group(self, target, label, output, subgroup):
|
97 |
+
""" Compute bias AUC metrics
|
98 |
+
"""
|
99 |
+
|
100 |
+
y_trues, y_preds = zip(*self._genreate_subgroup(target, label, output, subgroup))
|
101 |
+
subgroup_auc_score = roc_auc_score(y_trues, y_preds)
|
102 |
+
|
103 |
+
y_trues, y_preds = zip(*self._genreate_bpsn(target, label, output, subgroup))
|
104 |
+
bpsn_auc_score = roc_auc_score(y_trues, y_preds)
|
105 |
+
|
106 |
+
y_trues, y_preds = zip(*self._genreate_bnsp(target, label, output, subgroup))
|
107 |
+
bnsp_auc_score = roc_auc_score(y_trues, y_preds)
|
108 |
+
|
109 |
+
|
110 |
+
return {'Subgroup' : subgroup_auc_score,
|
111 |
+
'BPSN' : bpsn_auc_score,
|
112 |
+
'BNSP' : bnsp_auc_score}
|
113 |
+
|
114 |
+
def _update_overall(self, result, labels, outputs, power_value=-5):
|
115 |
+
"""Compute the generalized mean of Bias AUCs"""
|
116 |
+
result['Overall'] = {}
|
117 |
+
for metric in ['Subgroup', 'BPSN', 'BNSP']:
|
118 |
+
metric_values = np.array([result[community][metric] for community in result
|
119 |
+
if community != 'Overall'])
|
120 |
+
metric_values **= power_value
|
121 |
+
mean_value = np.power(np.sum(metric_values)/(len(result) - 1), 1/power_value)
|
122 |
+
result['Overall'][f"{metric} generalized mean"] = mean_value
|
123 |
+
y_preds = [output[1] for output in outputs]
|
124 |
+
result['Overall']["Overall AUC"] = roc_auc_score(labels, y_preds)
|
125 |
+
return result
|
126 |
+
|
127 |
+
|
128 |
+
def _compute(self, target, label, output, subgroups=None):
|
129 |
+
if subgroups is None:
|
130 |
+
subgroups = set(group for group_list in target for group in group_list)
|
131 |
+
result = {subgroup : self._auc_by_group(target, label, output, subgroup)
|
132 |
+
for subgroup in subgroups}
|
133 |
+
result = self._update_overall(result, label, output)
|
134 |
+
return result
|
135 |
+
|