-
Notifications
You must be signed in to change notification settings - Fork 13
/
evaluator.py
220 lines (195 loc) · 7.08 KB
/
evaluator.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
import os
from logging import getLogger
import numpy as np
from sklearn.metrics import roc_auc_score
from sklearn.metrics import log_loss
import pdb
class Evaluator:
def __init__(self, config):
self.logger = getLogger()
self.metric2func = {
'ndcg': self._calcu_nDCG,
'map': self._calcu_MAP,
'p': self._calcu_Precision,
'r': self._calcu_Recall,
'auc': roc_auc_score,
'logloss': log_loss,
}
self.topk = config['topk']
self.maxtopk = max(self.topk)
self.precision = config['metric_decimal_place']
self.metrics = ['r@5', 'p@5', 'ndcg@5', 'mrr']
self.base = []
self.idcg = []
for i in range(self.maxtopk):
self.base.append(np.log(2) / np.log(i + 2))
if i > 0:
self.idcg.append(self.base[i] + self.idcg[i - 1])
else:
self.idcg.append(self.base[i])
#self._load_geek2weak(config['dataset_path'])
def collect(self, interaction, scores, reverse=False):
uid2topk = {}
scores = scores.cpu().numpy()
labels = interaction['label'].numpy()
if reverse:
actor = 'job_id'
else:
actor = 'geek_id'
for i, uid in enumerate(interaction[actor].numpy()):
if uid not in uid2topk:
uid2topk[uid] = []
uid2topk[uid].append((scores[i], labels[i]))
return uid2topk
def evaluate(self, uid2topk_list, group='all'):
uid2topk = self._merge_uid2topk(uid2topk_list)
uid2topk = self._filter_illegal(uid2topk)
uid2topk = self._filter_group(uid2topk, group)
result = {}
result.update(self._calcu_ranking_metrics(uid2topk))
result.update(self._calcu_cls_metrics(uid2topk))
for m in result:
result[m] = round(result[m], self.precision)
return result, self._format_str(result)
def _format_str(self, result):
res = ''
for metric in self.metrics:
res += '\n\t{}:\t{:.4f}'.format(metric, result[metric])
return res
def _calcu_ranking_metrics(self, uid2topk):
result = {}
for m in ['ndcg', 'map', 'p', 'r']:
for k in self.topk:
metric = f'{m}@{k}'
if metric in self.metrics:
result[metric] = self.metric2func[m](uid2topk, k)
if 'mrr' in self.metrics:
result['mrr'] = self._calcu_MRR(uid2topk)
return result
def _calcu_cls_metrics(self, uid2topk):
scores, labels = self._flatten_cls_list(uid2topk)
result = {}
for m in ['auc', 'logloss']:
if m in self.metrics:
result[m] = self.metric2func[m](labels, scores)
if 'gauc' in self.metrics:
result['gauc'] = self._calcu_GAUC(uid2topk)
return result
def _calcu_GAUC(self, uid2topk):
weight_sum = auc_sum = 0
for uid, lst in uid2topk.items():
score_list, lb_list = zip(*lst)
scores = np.array(score_list)
labels = np.array(lb_list)
w = len(labels)
auc = roc_auc_score(labels, scores)
weight_sum += w
auc_sum += auc * w
return float(auc_sum / weight_sum)
def _calcu_nDCG(self, uid2topk, k):
tot = 0
for uid in uid2topk:
dcg = 0
pos = 0
for i, (score, lb) in enumerate(uid2topk[uid][:k]):
dcg += lb * self.base[i]
pos += lb
tot += dcg / self.idcg[int(pos) - 1]
return tot / len(uid2topk)
def _calcu_Precision(self, uid2topk, k):
tot = 0
for uid in uid2topk:
rec = 0
rel = 0
for i, (score, lb) in enumerate(uid2topk[uid][:k]):
rec += 1
rel += lb
tot += rel / rec
return tot / len(uid2topk)
def _calcu_Recall(self, uid2topk, k):
tot = 0
for uid in uid2topk:
rec = 0
rel = 0
for i, (score, lb) in enumerate(uid2topk[uid]):
if i < k:
rec += lb
rel += lb
tot += rec / rel
return tot / len(uid2topk)
def _calcu_MRR(self, uid2topk):
tot = 0
for uid in uid2topk:
for i, (score, lb) in enumerate(uid2topk[uid]):
if lb == 1:
tot += 1 / (i + 1)
break
return tot / len(uid2topk)
def _calcu_MAP(self, uid2topk, k):
tot = 0
for uid in uid2topk:
tp = 0
pos = 0
ap = 0
for i, (score, lb) in enumerate(uid2topk[uid][:k]):
if lb == 1:
tp += 1
pos += 1
ap += tp / (i + 1)
if pos > 0:
tot += ap / pos
return tot / len(uid2topk)
def _merge_uid2topk(self, uid2topk_list):
uid2topk = {}
for single_uid2topk in uid2topk_list:
for uid in single_uid2topk:
if uid not in uid2topk:
uid2topk[uid] = []
uid2topk[uid].extend(single_uid2topk[uid])
return self._sort_uid2topk(uid2topk)
def _load_geek2weak(self, dataset_path):
self.geek2weak = []
filepath = os.path.join(dataset_path, f'geek.weak')
self.logger.info(f'Loading {filepath}')
with open(filepath, 'r') as file:
for line in file:
token, weak = line.strip().split('\t')
self.geek2weak.append(int(weak))
def _filter_illegal(self, uid2topk):
new_uid2topk = {}
for uid, lst in uid2topk.items():
score_list, lb_list = zip(*lst)
lb_sum = sum(lb_list)
if lb_sum == 0 or lb_sum == len(lb_list):
continue
new_uid2topk[uid] = uid2topk[uid]
return new_uid2topk
def _filter_group(self, uid2topk, group):
if group == 'all':
return uid2topk
elif group in ['weak', 'skilled']:
self.logger.info(f'Evaluating on [{group}]')
flag = 1 if group == 'weak' else 0
new_uid2topk = {}
# for uid in uid2topk:
# if abs(self.geek2weak[uid] - flag) < 0.1:
# new_uid2topk[uid] = uid2topk[uid]
return new_uid2topk
else:
raise NotImplementedError(f'Not support [{group}]')
def _sort_uid2topk(self, uid2topk):
for uid in uid2topk:
uid2topk[uid].sort(key=lambda t: t[1], reverse=False)
uid2topk[uid].sort(key=lambda t: t[0], reverse=True)
return uid2topk
def _flatten_cls_list(self, uid2topk):
scores = []
labels = []
for uid, lst in uid2topk.items():
score_list, lb_list = zip(*lst)
scores.append(np.array(score_list))
labels.append(np.array(lb_list))
scores = np.concatenate(scores)
labels = np.concatenate(labels)
assert scores.shape[0] == labels.shape[0]
return scores, labels