-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathfml_disparate_impact_remover.py
246 lines (190 loc) · 11.8 KB
/
fml_disparate_impact_remover.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
from util.measure import performance_score, fariness_score
from util.io import save_dataframe, check_saving_path, save_model
from config.model import prep_construction, imb_construction, model_fml
from sklearn.model_selection import train_test_split
from aif360.datasets import StandardDataset
from aif360.algorithms.preprocessing import DisparateImpactRemover
from types import SimpleNamespace
import os
import csv
import json
import argparse
import numpy as np
import pandas as pd
def execute(cfg, mparam, grp):
''' Load settings '''
model_name = cfg.model_name
model_alg = mparam.model_alg
root_dir = cfg.root_dir
processed_dir = cfg.processed_dir
models_dir = cfg.models_dir
output_roc_dir = cfg.output_roc_dir
output_shap_dir = cfg.output_shap_dir
output_score_dir = cfg.output_score_dir
n_run = cfg.n_run
fairness_tab = pd.DataFrame(np.zeros((n_run, len(grp.fair_measure))))
fairness_tab.columns = grp.fair_measure
performance_tab = pd.DataFrame(np.zeros((n_run, len(grp.perf_measure))))
performance_tab.columns = grp.perf_measure
id_fairness_tab = pd.DataFrame(np.zeros((n_run, len(grp.fair_measure))))
id_fairness_tab.columns = grp.fair_measure
id_performance_tab = pd.DataFrame(np.zeros((n_run, len(grp.perf_measure))))
id_performance_tab.columns = grp.perf_measure
X = pd.read_csv(os.path.join(root_dir, processed_dir, cfg.input_features), index_col=0)
Y = pd.read_csv(os.path.join(root_dir, processed_dir, cfg.input_labels), index_col=0)
IX = pd.read_csv(os.path.join(root_dir, processed_dir, cfg.independent_full_features), index_col=0)
IY = pd.read_csv(os.path.join(root_dir, processed_dir, cfg.independent_labels), index_col=0)
for i in range(0,n_run):
print("==================================== iterate",i," running ========================")
random_seed = i
model_save_name = model_name + model_alg + grp.subgroup +"/"+ str(i)
#Generate Training and Testing Set
X_train, X_test, y_train, y_test = train_test_split(X, Y, stratify=Y, test_size=mparam.setting_params.train_test_ratio, random_state=random_seed)
#Generate Training and Evaluation Set
X_train, X_val, y_train, y_val = train_test_split(X_train, y_train, stratify=y_train, test_size=mparam.setting_params.train_val_ratio, random_state=random_seed) #0.125 * 0.8 = 0.1
''' Data Preprocessing'''
prep_c = prep_construction()
imbp = imb_construction(mparam)
X_train_features = prep_c.fit_transform(X_train)
X_test_features = prep_c.transform(X_test)
X_val_features = prep_c.transform(X_val)
IX_features = prep_c.fit_transform(IX)
if imbp is not None:
X_train_features, y_train = imbp.fit_resample(X_train_features, y_train)
X_train = pd.DataFrame(X_train_features, columns=X_train.columns)
X_test = pd.DataFrame(X_test_features, index = X_test.index, columns=X_test.columns)
X_val = pd.DataFrame(X_val_features, index = X_val.index, columns=X_val.columns)
IX_test = pd.DataFrame(IX_features, index = IX.index, columns=IX.columns)
''' Transform the Data to AIF-360 Data Format '''
train_df = pd.concat([X_train, y_train], axis=1, join='inner')
dataset_orig_train = StandardDataset(train_df,
grp.label_name,
grp.favorable_classes,
grp.protected_attribute_names,
grp.privileged_classes)
test_df = pd.concat([X_test, y_test], axis=1, join='inner')
dataset_orig_test = StandardDataset(test_df,
grp.label_name,
grp.favorable_classes,
grp.protected_attribute_names,
grp.privileged_classes)
val_df = pd.concat([X_val, y_val], axis=1, join='inner')
dataset_orig_valid = StandardDataset(val_df,
grp.label_name,
grp.favorable_classes,
grp.protected_attribute_names,
grp.privileged_classes)
independent_df = pd.concat([IX_test, IY], axis=1, join='inner')
dataset_orig_independent = StandardDataset(independent_df,
grp.label_name,
grp.favorable_classes,
grp.protected_attribute_names,
grp.privileged_classes)
'''
print(X_train_features.shape)
print(X_test_features.shape)
print(X_val_features.shape)
print(IX_features.shape)
print(X_train.shape)
print(X_test.shape)
print(X_val.shape)
print(IX_test.shape)
print(train_df.shape)
print(test_df.shape)
print(val_df.shape)
print(independent_df.shape)
exit()
'''
''' Pre-run Settings '''
unprivileged_groups = [{grp.protected_feature_name: grp.protected_feature_val}]
privileged_groups = [{grp.privileged_feature_name: grp.privileged_feature_val}]
protected_index = dataset_orig_train.feature_names.index(grp.protected_feature_name)
privileged_index = dataset_orig_train.feature_names.index(grp.privileged_feature_name)
tr_protected_features = np.reshape(dataset_orig_train.features[:,protected_index],[-1,1])
te_protected_features = np.reshape(dataset_orig_test.features[:,protected_index],[-1,1])
vl_protected_features = np.reshape(dataset_orig_valid.features[:,protected_index],[-1,1])
tr_protected_group_idx = dataset_orig_train.features[:,protected_index] == 1
tr_privileged_group_idx = dataset_orig_train.features[:,privileged_index] == 1
te_protected_group_idx = dataset_orig_test.features[:,protected_index] == 1
te_privileged_group_idx = dataset_orig_test.features[:,privileged_index] == 1
vl_protected_group_idx = dataset_orig_valid.features[:,protected_index] == 1
vl_privileged_group_idx = dataset_orig_valid.features[:,privileged_index] == 1
id_protected_group_idx = dataset_orig_independent.features[:,protected_index] == 1
id_privileged_group_idx = dataset_orig_independent.features[:,privileged_index] == 1
if (grp.is_mask_attr):
masked_index = []
masked_attrs = grp.masked_attrs
for attr in masked_attrs:
masked_index.append(dataset_orig_train.feature_names.index(attr) )
print(masked_index)
print(masked_attrs)
dataset_orig_train.features = np.delete(dataset_orig_train.features, masked_index, axis=1)
dataset_orig_test.features = np.delete(dataset_orig_test.features, masked_index, axis=1)
dataset_orig_valid.features = np.delete(dataset_orig_valid.features, masked_index, axis=1)
dataset_orig_independent.features = np.delete(dataset_orig_independent.features, masked_index, axis=1)
''' Data Pre-transforming'''
di = DisparateImpactRemover(repair_level=cfg.level)
train_repd = di.fit_transform(dataset_orig_train)
val_repd = di.fit_transform(dataset_orig_valid)
test_repd = di.fit_transform(dataset_orig_test)
independent_repd = di.fit_transform(dataset_orig_independent)
if (grp.is_mask_attr):
X_tr = train_repd.features
X_vl = val_repd.features
X_te = test_repd.features
X_id = independent_repd.features
else:
X_tr = np.delete(train_repd.features, [protected_index,privileged_index], axis=1)
X_vl = np.delete(val_repd.features, [protected_index,privileged_index], axis=1)
X_te = np.delete(test_repd.features, [protected_index,privileged_index], axis=1)
X_id = np.delete(independent_repd.features, [protected_index,privileged_index], axis=1)
y_tr = train_repd.labels.ravel()
y_vl = val_repd.labels.ravel()
y_id = independent_repd.labels.ravel()
print(X_tr.shape)
''' Run Modeling '''
clf, _, fit_params = model_fml(mparam, X_val=X_vl, y_val=y_vl)
clf.fit(X_tr, y_tr, **fit_params)
test_repd_pred = test_repd.copy()
test_repd_pred.labels = clf.predict(X_te)
test_repd_pred.scores = clf.predict_proba(X_te)
independent_repd_pred = independent_repd.copy()
independent_repd_pred.labels = clf.predict(X_id)
independent_repd_pred.scores = clf.predict_proba(X_id)
'''Prepare Results to the Output Format'''
y_protected_test = dataset_orig_test.labels[te_protected_group_idx]
y_privileged_test = dataset_orig_test.labels[te_privileged_group_idx]
y_protected_pred = test_repd_pred.labels[te_protected_group_idx]
y_privileged_pred = test_repd_pred.labels[te_privileged_group_idx]
y_protected_IY = dataset_orig_independent.labels[id_protected_group_idx]
y_privileged_IY = dataset_orig_independent.labels[id_privileged_group_idx]
y_protected_IY_pred = independent_repd_pred.labels[id_protected_group_idx]
y_privileged_IY_pred = independent_repd_pred.labels[id_privileged_group_idx]
fairness_tab.iloc[i] = fariness_score(y_protected_test, y_privileged_test, y_protected_pred, y_privileged_pred)
performance_tab.iloc[i] = performance_score(dataset_orig_test.labels, test_repd_pred.labels, test_repd_pred.scores[:, 1])
id_fairness_tab.iloc[i] = fariness_score(y_protected_IY, y_privileged_IY, y_protected_IY_pred, y_privileged_IY_pred)
id_performance_tab.iloc[i] = performance_score(dataset_orig_independent.labels, independent_repd_pred.labels, independent_repd_pred.scores[:, 1])
save_model(di, root_dir, models_dir,model_save_name,"di.pk")
save_model(clf, root_dir, models_dir,model_save_name,"clf.pk")
save_dataframe(fairness_tab, root_dir, output_score_dir, model_name+model_alg+grp.subgroup, "fairness.csv" )
save_dataframe(performance_tab, root_dir, output_score_dir, model_name+model_alg+grp.subgroup, "performance.csv" )
save_dataframe(id_fairness_tab, root_dir, output_score_dir, model_name+model_alg+grp.subgroup, "indenpendent_fairness.csv" )
save_dataframe(id_performance_tab, root_dir, output_score_dir, model_name+model_alg+grp.subgroup, "indenpendent_performance.csv" )
save_model({"config":cfg, "param":mparam, "group":grp}, root_dir, models_dir,model_name+model_alg+grp.subgroup,"experimental_config.pk")
print(fairness_tab)
print(performance_tab)
print(id_fairness_tab)
print(id_performance_tab)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--setting", "-s", type=str, required=True)
parser.add_argument("--model_params", "-m", type=str, required=True)
parser.add_argument("--group_info", "-g", type=str, required=True)
args = parser.parse_args()
with open(args.setting) as json_file:
cfg = json.load(json_file, object_hook=lambda d: SimpleNamespace(**d))
with open(args.model_params) as json_file:
mparam = json.load(json_file, object_hook=lambda d: SimpleNamespace(**d))
with open(args.group_info) as json_file:
grp = json.load(json_file, object_hook=lambda d: SimpleNamespace(**d))
execute(cfg, mparam, grp)