-
Notifications
You must be signed in to change notification settings - Fork 33
/
eval_metrics.py
196 lines (166 loc) · 8.67 KB
/
eval_metrics.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
import glob
import matplotlib.pyplot as plt
from sklearn.metrics import roc_auc_score, ndcg_score, log_loss, roc_curve
from collections import Counter, defaultdict
from scipy.stats import sem, t
import pandas as pd
import numpy as np
import json
import re
import os
import sys
SENSITIVE_ATTR = "school_id"
THIS_GROUP = 25
def avgstd(l):
'''
Given a list of values, returns a 95% confidence interval
if the standard deviation is unknown.
'''
n = len(l)
mean = sum(l) / n
if n == 1:
return '%.3f' % round(mean, 3)
std_err = sem(l)
confidence = 0.95
h = std_err * t.ppf((1 + confidence) / 2, n - 1)
return '%.3f ± %.3f' % (round(mean, 3), round(h, 3))
def all_metrics(results, test):
ndcg_ = defaultdict(list)
predictions_per_user = defaultdict(lambda: defaultdict(list))
predictions_per_sensitive_attr = defaultdict(lambda: defaultdict(list))
metrics_per_user = defaultdict(list)
metrics_per_sensitive_attr = defaultdict(list)
#roc_curves_per_sensitive_attr = defaultdict(lambda: defaultdict(list))
model = 'LR' if results['model'] == 'LR' else 'FM' + str(len(str(results['model'])))
print(model)
fold = results['predictions'][0]['fold']
y_pred = np.array(np.array(results['predictions'][0]['pred']))
y = np.array(results['predictions'][0]['y'])
try:
assert len(y) == len(test)
except AssertionError:
print('This is not the right fold', len(y), len(test))
sys.exit(0)
for user, pred, true in zip(test['user'], y_pred, y):
predictions_per_user[user]['pred'].append(pred)
predictions_per_user[user]['y'].append(true)
attribute = np.array(test[SENSITIVE_ATTR])
protected = np.argwhere(attribute % 2 == 0).reshape(-1)
unprotected = np.argwhere(attribute % 2 == 1).reshape(-1)
# protected = np.argwhere(attribute == THIS_GROUP).reshape(-1)
# unprotected = np.argwhere(attribute != THIS_GROUP).reshape(-1)
print(type(y))
print(len(y), len(y[protected]), len(y[unprotected]))
for attr, pred, true in zip(test[SENSITIVE_ATTR], y_pred, y):
predictions_per_sensitive_attr[attr]['pred'].append(pred)
predictions_per_sensitive_attr[attr]['y'].append(true)
users_ids = []
attr_ids = []
for user in predictions_per_user:
this_pred = np.array(predictions_per_user[user]['pred'])
this_true = np.array(predictions_per_user[user]['y'])
if len(this_pred) > 1:
users_ids.append(user)
# print(this_true)
metrics_per_user['nll'].append(log_loss(this_true, this_pred, labels=[0, 1]))
metrics_per_user['ndcg'].append(ndcg_score([this_true], [this_pred]))
metrics_per_user['ndcg@10'].append(ndcg_score([this_true], [this_pred], k=10))
metrics_per_user['ndcg-'].append(ndcg_score([1 - this_true], [1 - this_pred]))
ndcg_[model].append(ndcg_score([1 - this_true], [1 - this_pred]))
metrics_per_user['ndcg@10-'].append(ndcg_score([1 - this_true], [1 -this_pred], k=10))
if len(np.unique(this_true)) > 1:
metrics_per_user['auc'].append(roc_auc_score(this_true, this_pred))
nb_samples = []
for attr in predictions_per_sensitive_attr:
this_pred = np.array(predictions_per_sensitive_attr[attr]['pred'])
this_true = np.array(predictions_per_sensitive_attr[attr]['y'])
if len(this_pred) > 1:
metrics_per_sensitive_attr['ndcg'].append(ndcg_score([this_true], [this_pred]))
metrics_per_sensitive_attr['ndcg@10'].append(ndcg_score([this_true], [this_pred], k=10))
metrics_per_sensitive_attr['ndcg-'].append(ndcg_score([1 - this_true], [1 - this_pred]))
metrics_per_sensitive_attr['ndcg@10-'].append(ndcg_score([1 - this_true], [1 -this_pred], k=10))
if len(np.unique(this_true)) > 1:
attr_ids.append(attr)
nb_samples.append(len(this_true))
metrics_per_sensitive_attr['auc'].append(roc_auc_score(this_true, this_pred))
#roc_curves_per_sensitive_attr[attr]["fpr"], roc_curves_per_sensitive_attr[attr]["tpr"], _ = roc_curve(this_true, this_pred)
print('Test length', len(y))
print(y[:10], test[:10])
print('overall auc', np.round(roc_auc_score(y, y_pred), 3))
print('overall nll', np.round(log_loss(y, y_pred), 3))
print('sliced auc (per user)', avgstd(metrics_per_user['auc']))
print('sliced auc (per group)', avgstd(metrics_per_sensitive_attr['auc']))
print('sliced nll', avgstd(metrics_per_user['nll']))
"""
print('ndcg', avgstd(metrics_per_user['ndcg']))
print('ndcg@10', avgstd(metrics_per_user['ndcg@10']))
print('ndcg-', avgstd(metrics_per_user['ndcg-']))
print('ndcg@10-', avgstd(metrics_per_user['ndcg@10-']))
"""
# Display ids of the students that have the lowest/highest ndcg-
# print("Lowest NDCG- = {} on user {}".format(np.around(np.min(ndcg_[model]),5),users_ids[np.argmin(ndcg_[model])]))
# print(np.array(predictions_per_user[users_ids[np.argmin(ndcg_)]]['y']))
# print(np.array(predictions_per_user[users_ids[np.argmin(ndcg_)]]['pred']))
# print("Highest NDCG- = {} on user {}".format(np.around(np.max(ndcg_[model]),5),users_ids[np.argmax(ndcg_[model])]))
# print(np.array(predictions_per_user[users_ids[np.argmax(ndcg_)]]['y']))
# print(np.array(predictions_per_user[users_ids[np.argmax(ndcg_)]]['pred']))
"""
diff = abs(np.array(ndcg_[model]) - np.array(ndcg_['FM76']))
this_pos = np.argmax(diff)
this_user = users_ids[this_pos]
print("Biggest difference NDCG- = {} {}{} LR{} on user {}".format(np.max(diff), model, ndcg_[model][this_pos], ndcg_['FM76'][this_pos], this_user))
print(sorted(list(zip(predictions_per_user[this_user]['pred'], predictions_per_user[this_user]['y']))))
"""
candidates = Counter()
val = 0
for subgroup, auc, nb in zip(attr_ids, metrics_per_sensitive_attr['auc'], nb_samples):
candidates[subgroup] = (-auc, -nb)
print(len(candidates), 'groups and ', test[SENSITIVE_ATTR].nunique(), 'schools in test')
x = []
nb = []
for k, (xi, yi) in candidates.most_common():
if val < 5:
print(k, (-xi, -yi))
x.append(-xi)
nb.append(-yi)
val += 1
plt.stem(x, nb, use_line_collection=True)
plt.xlabel('AUC value')
plt.ylabel('Number of samples in group')
plt.title('For each group, number of samples per AUC value')
plt.show()
# Display ids of the subgroups (sensitive attribute) that have the lowest/highest AUC
print("Lowest AUC = {} on subgroup {}".format(np.around(np.min(metrics_per_sensitive_attr['auc']),5),
attr_ids[np.argmin(metrics_per_sensitive_attr['auc'])]))
# print(np.array(predictions_per_sensitive_attr[attr_ids[np.argmin(metrics_per_sensitive_attr['auc'])]]['y'])[:10])
# print(np.array(predictions_per_sensitive_attr[attr_ids[np.argmin(metrics_per_sensitive_attr['auc'])]]['pred'])[:10])
print("Highest AUC = {} on subgroup {}".format(np.around(np.max(metrics_per_sensitive_attr['auc']),5),
attr_ids[np.argmax(metrics_per_sensitive_attr['auc'])]))
# print(np.array(predictions_per_sensitive_attr[attr_ids[np.argmax(metrics_per_sensitive_attr['auc'])]]['y'])[:10])
# print(np.array(predictions_per_sensitive_attr[attr_ids[np.argmax(metrics_per_sensitive_attr['auc'])]]['pred'])[:10])
print('AUC of that group', roc_auc_score(y[protected], y_pred[protected]))
fpr_protec, tpr_protec, _ = roc_curve(y[protected], y_pred[protected])
# print('NDCG of that group', ndcg_score([y[protected]], [y_pred[protected]]))
print('AUC of other group', roc_auc_score(y[unprotected], y_pred[unprotected]))
# print('NDCG of other group', ndcg_score([y[unprotected]], [y_pred[unprotected]]))
fpr_unprotec, tpr_unprotec, _ = roc_curve(y[unprotected], y_pred[unprotected])
# Plot ROC curves of proteced vs. unprotected groups
plt.plot(fpr_protec, tpr_protec, label="Protected group")
plt.plot(fpr_unprotec, tpr_unprotec, label="Unprotected group")
plt.xlabel("False positive rate")
plt.ylabel("True positive rate")
plt.title("ROC curves comparison between protected and unprotected groups")
plt.legend()
plt.show()
if __name__ == '__main__':
os.chdir('data/assistments2009full')
# os.chdir('data/fr_en')
df = pd.read_csv('data.csv')
# r = re.compile(r'results-(.*).json')
# ndcg_ = defaultdict(list)
for filename in sorted(glob.glob('results*2024*'))[::-1][:1]:
print(filename)
with open(filename) as f:
results = json.load(f)
i_test = results['predictions'][0]['i_test']
all_metrics(results, df.iloc[i_test])