This repository has been archived by the owner on Dec 18, 2023. It is now read-only.
/
fairness.py
268 lines (228 loc) · 8.96 KB
/
fairness.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
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
from collections import defaultdict
from typing import List
import pandas as pd
from connect.evidence import MetricContainer, TableContainer
from credoai.artifacts import TabularData
from credoai.evaluators import Evaluator
from credoai.evaluators.utils.fairlearn import setup_metric_frames
from credoai.evaluators.utils.validation import (
check_artifact_for_nulls,
check_data_instance,
check_existence,
)
from credoai.modules.constants_metrics import (
MODEL_METRIC_CATEGORIES,
THRESHOLD_METRIC_CATEGORIES,
)
from credoai.modules.metrics import process_metrics
class ModelFairness(Evaluator):
"""
Model Fairness evaluator for Credo AI.
This evaluator calculates performance metrics disaggregated by a sensitive feature, as
well as evaluating the parity of those metrics.
Handles any metric that can be calculated on a set of ground truth labels and predictions,
e.g., binary classification, multiclass classification, regression.
Parameters
----------
metrics : List-like
list of metric names as string or list of Metrics (credoai.metrics.Metric).
Metric strings should in list returned by credoai.modules.list_metrics.
Note for performance parity metrics like
"false negative rate parity" just list "false negative rate". Parity metrics
are calculated automatically if the performance metric is supplied
method : str, optional
How to compute the differences: "between_groups" or "to_overall".
See fairlearn.metrics.MetricFrame.difference
for details, by default 'between_groups'
"""
required_artifacts = {"model", "data", "sensitive_feature"}
def __init__(
self,
metrics=None,
method="between_groups",
):
self.metrics = metrics
self.fairness_method = method
self.fairness_metrics = None
self.fairness_prob_metrics = None
super().__init__()
def _validate_arguments(self):
check_existence(self.metrics, "metrics")
check_data_instance(self.data, TabularData)
check_existence(self.data.sensitive_features, "sensitive_features")
check_artifact_for_nulls(self.data, "Data")
def _setup(self):
self.sensitive_features = self.data.sensitive_feature
self.y_true = self.data.y
self.y_pred = self.model.predict(self.data.X)
if hasattr(self.model, "predict_proba"):
self.y_prob = self.model.predict_proba(self.data.X)
else:
self.y_prob = (None,)
self.update_metrics(self.metrics)
def evaluate(self):
"""
Run fairness base module.
"""
fairness_results = self.get_fairness_results()
disaggregated_metrics = self.get_disaggregated_performance()
disaggregated_thresh_results = self.get_disaggregated_threshold_performance()
results = []
for result_obj in [
fairness_results,
disaggregated_metrics,
disaggregated_thresh_results,
]:
if result_obj is not None:
try:
results += result_obj
except TypeError:
results.append(result_obj)
self.results = results
return self
def update_metrics(self, metrics, replace=True):
"""
Replace metrics
Parameters
----------
metrics : List-like
list of metric names as string or list of Metrics (credoai.metrics.Metric).
Metric strings should in list returned by credoai.modules.list_metrics.
Note for performance parity metrics like
"false negative rate parity" just list "false negative rate". Parity metrics
are calculated automatically if the performance metric is supplied
"""
if replace:
self.metrics = metrics
else:
self.metrics += metrics
self.processed_metrics, self.fairness_metrics = process_metrics(
self.metrics, self.model.type
)
self.metric_frames = setup_metric_frames(
self.processed_metrics,
self.y_pred,
self.y_prob,
self.y_true,
self.sensitive_features,
)
def get_disaggregated_performance(self):
"""
Return performance metrics for each group
Parameters
----------
melt : bool, optional
If True, return a long-form dataframe, by default False
Returns
-------
TableContainer
The disaggregated performance metrics
"""
disaggregated_df = pd.DataFrame()
for name, metric_frame in self.metric_frames.items():
if name == "thresh":
continue
df = metric_frame.by_group.copy().convert_dtypes()
disaggregated_df = pd.concat([disaggregated_df, df], axis=1)
if disaggregated_df.empty:
self.logger.warn("No disaggregated metrics could be calculated.")
return
# reshape
disaggregated_results = disaggregated_df.reset_index().melt(
id_vars=[disaggregated_df.index.name],
var_name="type",
)
disaggregated_results.name = "disaggregated_performance"
metric_type_label = {
"metric_types": disaggregated_results.type.unique().tolist()
}
return TableContainer(
disaggregated_results,
**self.get_info(labels=metric_type_label),
)
def get_disaggregated_threshold_performance(self):
"""
Return performance metrics for each group
Parameters
----------
melt : bool, optional
If True, return a long-form dataframe, by default False
Returns
-------
List[TableContainer]
The disaggregated performance metrics
"""
metric_frame = self.metric_frames.get("thresh")
if metric_frame is None:
return
df = metric_frame.by_group.copy().convert_dtypes()
df = df.reset_index().melt(
id_vars=[df.index.name],
var_name="type",
)
to_return = defaultdict(list)
for i, row in df.iterrows():
tmp_df = row["value"]
tmp_df = tmp_df.assign(**row.drop("value"))
to_return[row["type"]].append(tmp_df)
for key in to_return.keys():
df = pd.concat(to_return[key])
df.name = "threshold_dependent_disaggregated_performance"
to_return[key] = df
disaggregated_thresh_results = []
for key, df in to_return.items():
labels = {"metric_type": key}
disaggregated_thresh_results.append(
TableContainer(df, **self.get_info(labels=labels))
)
return disaggregated_thresh_results
def get_fairness_results(self):
"""Return fairness and performance parity metrics
Note, performance parity metrics are labeled with their
related performance label, but are computed using
fairlearn.metrics.MetricFrame.difference(method)
Returns
-------
MetricContainer
The returned fairness metrics
"""
results = []
for metric_name, metric in self.fairness_metrics.items():
pred_argument = {"y_pred": self.y_pred}
if metric.takes_prob:
pred_argument = {"y_prob": self.y_prob}
try:
metric_value = metric.fun(
y_true=self.y_true,
sensitive_features=self.sensitive_features,
method=self.fairness_method,
**pred_argument,
)
except Exception as e:
self.logger.error(
f"A metric ({metric_name}) failed to run. "
"Are you sure it works with this kind of model and target?\n"
)
raise e
results.append({"metric_type": metric_name, "value": metric_value})
results = pd.DataFrame.from_dict(results)
# add parity results
parity_results = pd.Series(dtype=float)
parity_results = []
for name, metric_frame in self.metric_frames.items():
if name == "thresh":
# Don't calculate difference for curve metrics. This is not mathematically well-defined.
continue
diffs = metric_frame.difference(self.fairness_method).rename(
"{}_parity".format
)
diffs = pd.DataFrame({"metric_type": diffs.index, "value": diffs.values})
parity_results.append(diffs)
if parity_results:
parity_results = pd.concat(parity_results)
results = pd.concat([results, parity_results])
results.rename({"metric_type": "type"}, axis=1, inplace=True)
if results.empty:
self.logger.info("No fairness metrics calculated.")
return
return MetricContainer(results, **self.get_info())