Spaces:
Runtime error
Runtime error
File size: 14,584 Bytes
d2a8669 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 |
import numpy as np
import pandas as pd
class Fairness(object):
"""
Compute fairness metrics
"""
def __init__(self, G, test_nodes_idx, targets, predictions, sens_attr, neptune_run,
multiclass_pred=False, multiclass_sens=False):
self.multiclass_pred = multiclass_pred
self.multiclass_sens = multiclass_sens
self.sens_attr = sens_attr
self.neptune_run = neptune_run
self.neptune_run["sens_attr"] = self.sens_attr
self.G = G
self.test_nodes_idx = test_nodes_idx.cpu().detach().numpy()
self.true_y = np.asarray(targets) # target variables
self.pred_y = np.asarray(predictions) # prediction of the classifier
self.sens_attr_array = self.G.nodes["user"].data[self.sens_attr].cpu().detach().numpy() # sensitive attribute values
self.sens_attr_values = self.sens_attr_array[self.test_nodes_idx]
if self.multiclass_pred and self.multiclass_sens: # Classifier: multiclass - Sens.attr: multiclass
self.class_range = list(set(self.true_y))
self.y_hat = []
self.yneq_hat = []
for y_hat_idx in self.class_range:
self.y_hat.append(self.pred_y == y_hat_idx)
self.yneq_hat.append(self.pred_y != y_hat_idx)
self.sens_attr_range = list(set(self.sens_attr_values))
self.s = []
for s_idx in self.sens_attr_range:
self.s.append(self.sens_attr_values == s_idx)
self.y_s = []
self.yneq_s = []
for y_idx in self.class_range:
self.y_s.append([])
self.yneq_s.append([])
for s_idx in self.sens_attr_range:
self.y_s[y_idx].append(np.bitwise_and(self.true_y == y_idx, self.s[s_idx]))
self.yneq_s[y_idx].append(np.bitwise_and(self.true_y != y_idx, self.s[s_idx]))
self.y_s = np.array(self.y_s)
self.yneq_s = np.array(self.yneq_s)
elif self.multiclass_sens: # Classifier: binary - Sens.attr: multiclass
self.sens_attr_range = list(set(self.sens_attr_values))
self.s = []
self.y1_s = []
self.y0_s = []
for s_idx in self.sens_attr_range:
self.s.append(self.sens_attr_values == s_idx)
self.y1_s.append(np.bitwise_and(self.true_y == 1, self.s[s_idx]))
self.y0_s.append(np.bitwise_and(self.true_y == 0, self.s[s_idx]))
else: # Classifier: binary - Sens.attr: binary
self.s0 = self.sens_attr_values == 0
self.s1 = self.sens_attr_values == 1
self.y1_s0 = np.bitwise_and(self.true_y == 1, self.s0)
self.y1_s1 = np.bitwise_and(self.true_y == 1, self.s1)
self.y0_s0 = np.bitwise_and(self.true_y == 0, self.s0)
self.y0_s1 = np.bitwise_and(self.true_y == 0, self.s1)
def statistical_parity(self):
if self.multiclass_pred and self.multiclass_sens: # Classifier: multiclass - Sens.attr: multiclass
"""
P(y^=0|s=0) = P(y^=0|s=1) = ... = P(y^=0|s=N)
[...]
P(y^=M|s=0) = P(y^=M|s=1) = ... = P(y^=M|s=N)
"""
stat_parity = []
for y_hat_idx in self.class_range:
stat_parity.append([])
for s_idx in self.sens_attr_range:
stat_parity[y_hat_idx].append(
sum(np.bitwise_and(self.y_hat[y_hat_idx], self.s[s_idx])) /
sum(self.s[s_idx])
)
self.neptune_run["fairness/SP_y^" + str(y_hat_idx) + "_s" + str(s_idx)] = stat_parity[y_hat_idx][s_idx]
elif self.multiclass_sens: # Classifier: binary - Sens.attr: multiclass
''' P(y^=1|s=0) = P(y^=1|s=1) = ... = P(y^=1|s=N) '''
stat_parity_s = []
for s_idx in self.sens_attr_range:
stat_parity_s.append(sum(self.pred_y[self.s[s_idx]]) / sum(self.s[s_idx]))
self.neptune_run["fairness/SP_s" + str(s_idx)] = stat_parity_s[s_idx]
else: # Classifier: binary - Sens.attr: binary
''' P(y^=1|s=0) = P(y^=1|s=1) '''
# stat_parity = abs(sum(self.pred_y[self.s0]) / sum(self.s0) - sum(self.pred_y[self.s1]) / sum(self.s1))
stat_parity_s0 = sum(self.pred_y[self.s0]) / sum(self.s0)
stat_parity_s1 = sum(self.pred_y[self.s1]) / sum(self.s1)
stat_parity_diff = stat_parity_s0 - stat_parity_s1
self.neptune_run["fairness/SP_s0"] = stat_parity_s0
self.neptune_run["fairness/SP_s1"] = stat_parity_s1
print("Statistical Parity Difference (SPD): {:.4f}".format(np.abs(stat_parity_diff)))
self.neptune_run["fairness/SPD"] = stat_parity_diff
def equal_opportunity(self):
if self.multiclass_pred and self.multiclass_sens: # Classifier: multiclass - Sens.attr: multiclass
"""
P(y^=0|y=0,s=0) = P(y^=0|y=0,s=1) = ... = P(y^=0|y=0,s=N)
[...]
P(y^=M|y=M,s=0) = P(y^=M|y=M,s=1) = ... = P(y^=M|y=M,s=N)
"""
equal_opp = []
for y_hat_idx in self.class_range:
equal_opp.append([])
for s_idx in self.sens_attr_range:
try:
equal_opp[y_hat_idx].append(
sum(np.bitwise_and(self.y_hat[y_hat_idx], self.y_s[y_hat_idx][s_idx])) /
sum(self.y_s[y_hat_idx][s_idx])
)
except ZeroDivisionError:
equal_opp[y_hat_idx].append(0)
self.neptune_run["fairness/EO_y" + str(y_hat_idx) + "_s" + str(s_idx)] = equal_opp[y_hat_idx][s_idx]
elif self.multiclass_sens: # Classifier: binary - Sens.attr: multiclass
''' P(y^=1|y=1,s=0) = P(y^=1|y=1,s=1) = ... = P(y^=1|y=1,s=N) '''
equal_opp_s = []
for s_idx in self.sens_attr_range:
equal_opp_s.append(sum(self.pred_y[self.y1_s[s_idx]]) / sum(self.y1_s[s_idx]))
self.neptune_run["fairness/EO_s" + str(s_idx)] = equal_opp_s[s_idx]
else: # Classifier: binary - Sens.attr: binary
''' P(y^=1|y=1,s=0) = P(y^=1|y=1,s=1) '''
# equal_opp = abs(sum(self.pred_y[self.y1_s0]) / sum(self.y1_s0) - sum(self.pred_y[self.y1_s1]) / sum(self.y1_s1))
equal_opp_s0 = sum(self.pred_y[self.y1_s0]) / sum(self.y1_s0)
equal_opp_s1 = sum(self.pred_y[self.y1_s1]) / sum(self.y1_s1)
equal_opp_diff = equal_opp_s0 - equal_opp_s1
self.neptune_run["fairness/EO_s0"] = equal_opp_s0
self.neptune_run["fairness/EO_s1"] = equal_opp_s1
print("Equal Opportunity Difference (EOD): {:.4f}".format(np.abs(equal_opp_diff)))
self.neptune_run["fairness/EOD"] = equal_opp_diff
def overall_accuracy_equality(self):
if self.multiclass_pred and self.multiclass_sens: # Classifier: multiclass - Sens.attr: multiclass
''' P(y^=0|y=0,s=0) + ... + P(y^=M|y=M,s=0) = ... = P(y^=0|y=0,s=N) + ... + P(y^=M|y=M,s=N) '''
oae_s = []
for s_idx in self.sens_attr_range:
oae_temp = 0.0
for y_hat_idx in self.class_range:
try:
oae_temp += (
sum(np.bitwise_and(self.y_hat[y_hat_idx], self.y_s[y_hat_idx][s_idx])) /
sum(self.y_s[y_hat_idx][s_idx])
)
except ZeroDivisionError:
oae_temp += 0.0
oae_s.append(oae_temp)
self.neptune_run["fairness/OAE_s" + str(s_idx)] = oae_s[s_idx]
elif self.multiclass_sens: # Classifier: binary - Sens.attr: multiclass
''' P(y^=0|y=0,s=0) + P(y^=1|y=1,s=0) = ... = P(y^=0|y=0,s=N) + P(y^=1|y=1,s=N)'''
oae_s = []
for s_idx in self.sens_attr_range:
oae_s.append(
np.count_nonzero(self.pred_y[self.y0_s[s_idx]]==0) / sum(self.y0_s[s_idx]) +
sum(self.pred_y[self.y1_s[s_idx]]) / sum(self.y1_s[s_idx])
)
self.neptune_run["fairness/OAE_s" + str(s_idx)] = oae_s[s_idx]
else: # Classifier: binary - Sens.attr: binary
''' P(y^=0|y=0,s=0) + P(y^=1|y=1,s=0) = P(y^=0|y=0,s=1) + P(y^=1|y=1,s=1) '''
oae_s0 = np.count_nonzero(self.pred_y[self.y0_s0]==0) / sum(self.y0_s0) + sum(self.pred_y[self.y1_s0]) / sum(self.y1_s0)
oae_s1 = np.count_nonzero(self.pred_y[self.y0_s1]==0) / sum(self.y0_s1) + sum(self.pred_y[self.y1_s1]) / sum(self.y1_s1)
# oae_diff = abs(oae_s0 - oae_s1)
oae_diff = oae_s0 - oae_s1
self.neptune_run["fairness/OAE_s0"] = oae_s0
self.neptune_run["fairness/OAE_s1"] = oae_s1
print("Overall Accuracy Equality Difference (OAED): {:.4f}".format(np.abs(oae_diff)))
self.neptune_run["fairness/OAED"] = oae_diff
def treatment_equality(self):
if self.multiclass_pred and self.multiclass_sens: # Classifier: multiclass - Sens.attr: multiclass
"""
P(y^=0|y/=0,s=0) / P(y^/=0|y=0,s=0) = ... = P(y^=0|y/=0,s=N) / P(y^/=0|y=0,s=N)
[...]
P(y^=M|y/=M,s=0) / P(y^/=M|y=M,s=0) = ... = P(y^=M|y/=M,s=N) / P(y^/M|y=M,s=N)
"""
te_fp_fn = []
te_fn_fp = []
te = []
for y_hat_idx in self.class_range:
te_fp_fn.append([])
te_fn_fp.append([])
abs_te_fp_fn = 0.0
abs_te_fn_fp = 0.0
te.append([])
for s_idx in self.sens_attr_range:
try:
te_fp_fn[y_hat_idx].append(
(sum(np.bitwise_and(self.y_hat[y_hat_idx], self.yneq_s[y_hat_idx][s_idx])) / sum(self.yneq_s[y_hat_idx][s_idx])) /
(sum(np.bitwise_and(self.yneq_hat[y_hat_idx], self.y_s[y_hat_idx][s_idx])) / sum(self.y_s[y_hat_idx][s_idx]))
)
except ZeroDivisionError:
te_fp_fn[y_hat_idx].append(0)
try:
te_fn_fp[y_hat_idx].append(
(sum(np.bitwise_and(self.yneq_hat[y_hat_idx], self.y_s[y_hat_idx][s_idx])) / sum(self.y_s[y_hat_idx][s_idx])) /
(sum(np.bitwise_and(self.y_hat[y_hat_idx], self.yneq_s[y_hat_idx][s_idx])) / sum(self.yneq_s[y_hat_idx][s_idx]))
)
except ZeroDivisionError:
te_fn_fp[y_hat_idx].append(0)
abs_te_fp_fn += abs(te_fp_fn[y_hat_idx][s_idx])
abs_te_fn_fp += abs(te_fn_fp[y_hat_idx][s_idx])
if abs_te_fp_fn < abs_te_fn_fp:
te[y_hat_idx].append(te_fp_fn[y_hat_idx][s_idx])
else:
te[y_hat_idx].append(te_fn_fp[y_hat_idx][s_idx])
for y_idx in self.class_range:
for s_idx in self.sens_attr_range:
self.neptune_run["fairness/TE_y" + str(y_idx) + "_s" + str(s_idx)] = te[y_idx][s_idx]
elif self.multiclass_sens: # Classifier: binary - Sens.attr: multiclass
''' P(y^=1|y=0,s=0) / P(y^=0|y=1,s=0) = ... = P(y^=1|y=0,s=N) / P(y^=0|y=1,s=N) '''
te1_s = []
te0_s = []
abs_te1 = []
abs_te0 = []
for s_idx in self.sens_attr_range:
te1_s.append(
(sum(self.pred_y[self.y0_s[s_idx]]) / sum(self.y0_s[s_idx])) /
(np.count_nonzero(self.pred_y[self.y1_s[s_idx]]==0) / sum(self.y1_s[s_idx]))
)
te0_s.append(
(np.count_nonzero(self.pred_y[self.y1_s[s_idx]]==0) / sum(self.y1_s[s_idx])) /
(sum(self.pred_y[self.y0_s[s_idx]]) / sum(self.y0_s[s_idx]))
)
abs_te1.append(abs(te1_s[s_idx]))
abs_te0.append(abs(te0_s[s_idx]))
if sum(abs_te1) < sum(abs_te0):
te_s = te1_s
else:
te_s = te0_s
#for i in self.sens_attr_range:
self.neptune_run["fairness/TE_s" + str(i)] = te_s[i]
else: # Classifier: binary - Sens.attr: binary
''' P(y^=1|y=0,s=0) / P(y^=0|y=1,s=0) = P(y^=1|y=0,s=1) / P(y^=0|y=1,s=1) '''
te1_s0 = (sum(self.pred_y[self.y0_s0]) / sum(self.y0_s0)) / (np.count_nonzero(self.pred_y[self.y1_s0]==0) / sum(self.y1_s0))
te1_s1 = (sum(self.pred_y[self.y0_s1]) / sum(self.y0_s1)) / (np.count_nonzero(self.pred_y[self.y1_s1]==0) / sum(self.y1_s1))
te_diff_1 = te1_s0 - te1_s1
abs_ted_1 = abs(te_diff_1)
''' P(y^=0|y=1,s=0) / P(y^=1|y=0,s=0) = P(y^=0|y=1,s=1) / P(y^=1|y=0,s=1) '''
te0_s0 = (np.count_nonzero(self.pred_y[self.y1_s0]==0) / sum(self.y1_s0)) / (sum(self.pred_y[self.y0_s0]) / sum(self.y0_s0))
te0_s1 = (np.count_nonzero(self.pred_y[self.y1_s1]==0) / sum(self.y1_s1)) / (sum(self.pred_y[self.y0_s1]) / sum(self.y0_s1))
te_diff_0 = te0_s0 - te0_s1
abs_ted_0 = abs(te_diff_0)
# te_diff = min(te_diff_1, te_diff_0)
if abs_ted_0 < abs_ted_1:
te_s0 = te0_s0
te_s1 = te0_s1
te_diff = te_diff_0
else:
te_s0 = te1_s0
te_s1 = te1_s1
te_diff = te_diff_1
self.neptune_run["fairness/TE_s0"] = te_s0
self.neptune_run["fairness/TE_s1"] = te_s1
print("Treatment Equality Difference (TED): {:.4f}".format(np.abs(te_diff)))
self.neptune_run["fairness/TED"] = te_diff
def disparate_impact(self):
#num_of_priv = sum(self.s0)
#num_of_unpriv = sum(self.s1)
#unpriv_ratio = sum(self.pred_y[self.]/num_of_unpriv
#stat_parity_s0 = sum(self.pred_y[self.s0]) / sum(self.s0)
#stat_parity_s1 = sum(self.pred_y[self.s1]) / sum(self.s1)
#stat_parity_diff = stat_parity_s0 - stat_parity_s1
print('true_y:', self.true_y)
print('pred_y:', self.pred_y) |