-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
324 lines (288 loc) · 12.2 KB
/
utils.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
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
import collections
import math
import os
import pdb
import numpy as np
import scipy
import torch
from scipy import stats
from tqdm import tqdm
import torch.nn.functional as F
TOP_Ks = [20, 50, 100]
def RecallPrecision_ATk(test_data, r, k):
"""
test_data should be a list? cause users may have different amount of pos items. shape (test_batch, k)
pred_data : shape (test_batch, k) NOTE: pred_data should be pre-sorted
k : top-k
"""
right_pred = r[:, :k].sum(1)
precis_n = k
recall_n = np.array([len(test_data[i]) for i in range(len(test_data))])
recall = np.sum(right_pred / recall_n)
precis = np.sum(right_pred) / precis_n
return {'recall': recall, 'precision': precis}
def RecallPrecision_ATk_cointer(test_data, r, k):
"""
test_data should be a list? cause users may have different amount of pos items. shape (test_batch, k)
pred_data : shape (test_batch, k) NOTE: pred_data should be pre-sorted
k : top-k
"""
right_pred = r[:, :k].sum(1)
precis_n = k
recall_n = np.array([len(test_data[i]) for i in range(len(test_data))])
recall = right_pred / recall_n
precis = right_pred / precis_n
return {'recall': recall, 'precision': precis}
def NDCGatK_r_cointer(test_data, r, k):
assert len(r) == len(test_data)
pred_data = r[:, :k]
test_matrix = np.zeros((len(pred_data), k))
for i, items in enumerate(test_data):
length = k if k <= len(items) else len(items)
test_matrix[i, :length] = 1
max_r = test_matrix
idcg = np.sum(max_r * 1. / np.log2(np.arange(2, k + 2)), axis=1)
dcg = pred_data * (1. / np.log2(np.arange(2, k + 2)))
dcg = np.sum(dcg, axis=1)
idcg[idcg == 0.] = 1.
ndcg = dcg / idcg
ndcg[np.isnan(ndcg)] = 0.
return ndcg
def NDCGatK_r(test_data, r, k):
"""
Normalized Discounted Cumulative Gain
rel_i = 1 or 0, so 2^{rel_i} - 1 = 1 or 0
"""
assert len(r) == len(test_data)
pred_data = r[:, :k]
test_matrix = np.zeros((len(pred_data), k))
for i, items in enumerate(test_data):
length = k if k <= len(items) else len(items)
test_matrix[i, :length] = 1
max_r = test_matrix
idcg = np.sum(max_r * 1. / np.log2(np.arange(2, k + 2)), axis=1)
dcg = pred_data * (1. / np.log2(np.arange(2, k + 2)))
dcg = np.sum(dcg, axis=1)
idcg[idcg == 0.] = 1.
ndcg = dcg / idcg
ndcg[np.isnan(ndcg)] = 0.
return np.sum(ndcg)
def minibatch(*tensors, **kwargs):
batch_size = kwargs.get('batch_size', 2048)
if len(tensors) == 1:
tensor = tensors[0]
for i in range(0, len(tensor), batch_size):
yield tensor[i:i + batch_size]
else:
for i in range(0, len(tensors[0]), batch_size):
yield tuple(x[i:i + batch_size] for x in tensors)
def getLabel(test_data, pred_data):
r = []
for i in range(len(test_data)):
groundTrue = test_data[i]
predictTopK = pred_data[i]
pred = list(map(lambda x: x in groundTrue, predictTopK))
pred = np.array(pred).astype("float")
r.append(pred)
return np.array(r).astype('float')
def test_one_batch(X, topks, global_pop=None):
sorted_items = X[0].numpy()
groundTrue = X[1]
r = getLabel(groundTrue, sorted_items)
pre, recall, ndcg, arp = [], [], [], []
for k in topks:
ret = RecallPrecision_ATk_cointer(groundTrue, r, k)
pre.append(ret['precision'])
recall.append(ret['recall'])
ndcg.append(NDCGatK_r_cointer(groundTrue, r, k))
if global_pop is not None:
arp.append(global_pop[sorted_items[:, :k]].mean(1).tolist())
ret = {'recall': recall,
'precision': pre,
'ndcg': ndcg}
if global_pop is not None:
ret['arp'] = arp
return ret
def test_one_batch_for_global_novelty(rating, topks, global_nov, global_pop):
nov = dict()
pru = dict()
for k in topks:
topk_items = rating[:, :k]
nov[k] = global_nov[topk_items].mean(1).tolist()
pru[k] = [-stats.spearmanr(global_pop[topk_items][i], torch.arange(k))[0] for i in range(len(topk_items))]
return nov, pru
def test_one_batch_local_novelty(rating, topks, local_nov, batch_users, local_pop):
nov = dict()
pru = dict()
for k in topks:
topk_items = rating[:, :k]
nov[k] = torch.gather(local_nov[batch_users], 1, topk_items).mean(1).tolist()
topk_pop = torch.gather(local_pop[batch_users], 1, topk_items)
pru[k] = []
for i in range(len(topk_items)):
src = -stats.spearmanr(topk_pop[i], torch.arange(k))[0]
if math.isnan(src):
src = 0
pru[k].append(src)
return nov, pru
def cal_global_nov(train_records, num_items):
pop = [0] * num_items
nov = [1] * num_items
for user in train_records:
for item in train_records[user]:
pop[item] += 1
u_num = len(train_records)
for i in range(num_items):
if pop[i] > 0:
nov[i] = -(1 / np.log2(u_num)) * np.log2(pop[i] / u_num)
return torch.tensor(nov), torch.tensor(pop)
def cal_local_nov(dataset, sim_users, train_records, num_items):
sim_coe = len(max(sim_users.values(), key=len))
local_nov, local_pop = None, None
if os.path.exists(f'co_items/local_nov-{dataset}-{sim_coe}.npy'):
local_nov = np.load(f'co_items/local_nov-{dataset}-{sim_coe}.npy', allow_pickle=True)
if os.path.exists(f'co_items/local_pop-{dataset}-{sim_coe}.npy'):
local_pop = np.load(f'co_items/local_pop-{dataset}-{sim_coe}.npy', allow_pickle=True)
if local_nov is None or local_pop is None:
u_num = max(train_records.keys()) + 1
local_nov = [[1] * num_items for _ in range(u_num)]
local_pop = [[0] * num_items for _ in range(u_num)]
train_item_records = collections.defaultdict(set)
for user in train_records:
for item in train_records[user]:
train_item_records[item].add(user)
for item in tqdm(train_item_records):
for user in range(u_num):
di = len(train_item_records[item].intersection(sim_users[user]))
if di != 0:
local_pop[user][item] = di
local_nov[user][item] = -(1 / np.log2(len(sim_users[user]))) * np.log2(di / len(sim_users[user]))
np.save(f'co_items/local_nov-{dataset}-{sim_coe}.npy', local_nov)
np.save(f'co_items/local_pop-{dataset}-{sim_coe}.npy', local_pop)
return torch.tensor(local_nov), torch.tensor(local_pop)
def test_model(model, train_records, test_records, global_pop, graph=None):
model.eval()
max_K = max(TOP_Ks)
results = {'precision': dict(),
'recall': dict(),
'ndcg': dict(),
'arp': dict()}
for topk in TOP_Ks:
results['precision'][topk] = []
results['recall'][topk] = []
results['ndcg'][topk] = []
results['arp'][topk] = []
with torch.no_grad():
users = list(test_records.keys())
users_list = []
rating_list = []
groundTrue_list = []
u_batch_size = 128 if hasattr(model, 'sys_params') and 'ncf' in model.sys_params.model else 8192
total_batch = len(users) // u_batch_size + 1
for batch_users in minibatch(users, batch_size=u_batch_size):
users_list.append(batch_users)
allPos = [train_records[u] for u in batch_users]
groundTrue = [test_records[u] for u in batch_users]
batch_users_gpu = torch.Tensor(batch_users).long()
batch_users_gpu = batch_users_gpu.to(model.device)
if graph:
rating = model.get_user_ratings(batch_users_gpu, graph)
else:
rating = model.get_user_ratings(batch_users_gpu)
exclude_index = []
exclude_items = []
for range_i, items in enumerate(allPos):
exclude_index.extend([range_i] * len(items))
exclude_items.extend(items)
rating[exclude_index, exclude_items] = -float('inf')
_, rating_K = torch.topk(rating, k=max_K)
rating_list.append(rating_K.cpu())
groundTrue_list.append(groundTrue)
assert total_batch == len(users_list)
X = zip(rating_list, groundTrue_list)
pre_results = []
for x in X:
pre_results.append(test_one_batch(x, TOP_Ks, global_pop))
scale = float(u_batch_size / len(users))
for result in pre_results:
for idx, topk in enumerate(TOP_Ks):
results['recall'][topk].extend(result['recall'][idx])
results['precision'][topk].extend(result['precision'][idx])
results['ndcg'][topk].extend(result['ndcg'][idx])
results['arp'][topk].extend(result['arp'][idx])
results['users'] = [user for user_lst in users_list for user in user_lst]
return results
def test_cointer_model(model, train_records, sim_users, test_records, i_num, graph=None):
model.eval()
max_K = max(TOP_Ks)
with torch.no_grad():
users = list(test_records.keys())
users_list = []
rating_list = []
batch_users_list = []
u_batch_size = 128 # 8192
total_batch = len(users) // u_batch_size + 1
for batch_users in minibatch(users, batch_size=u_batch_size):
users_list.append(batch_users)
allPos = [train_records[u] for u in batch_users]
batch_users_gpu = torch.Tensor(batch_users).long()
batch_users_gpu = batch_users_gpu.to(model.device)
if graph:
rating = model.get_user_ratings(batch_users_gpu, graph)
else:
rating = model.get_user_ratings(batch_users_gpu)
exclude_index = []
exclude_items = []
for range_i, items in enumerate(allPos):
exclude_index.extend([range_i] * len(items))
exclude_items.extend(items)
rating[exclude_index, exclude_items] = 0
_, rating_K = torch.topk(rating, k=max_K)
rating_list.append(rating_K.cpu())
batch_users_list.append(batch_users)
assert total_batch == len(users_list)
X = zip(rating_list, batch_users_list)
GNov_results = collections.defaultdict(list)
LNov_results = collections.defaultdict(list)
GPRU_results = collections.defaultdict(list)
LPRU_results = collections.defaultdict(list)
global_nov, global_pop = cal_global_nov(train_records, i_num)
local_nov, local_pop = cal_local_nov(model.dataset, sim_users, train_records, i_num)
for rating, batch_users in X:
glo_nov, glo_pru = test_one_batch_for_global_novelty(rating, TOP_Ks, global_nov, global_pop)
loc_nov, loc_pru = test_one_batch_local_novelty(rating, TOP_Ks, local_nov, batch_users, local_pop)
for topk in TOP_Ks:
GNov_results[topk].extend(glo_nov[topk])
LNov_results[topk].extend(loc_nov[topk])
GPRU_results[topk].extend(glo_pru[topk])
LPRU_results[topk].extend(loc_pru[topk])
return GNov_results, LNov_results, GPRU_results, LPRU_results, global_pop, local_pop
def clear_test_file():
# dataset = 'TaobaoAd'
# domains = ['d0', 'd1', 'd2', 'd3', 'd4', 'd5', 'd6']
dataset = 'ml-1M'
domains = ['Arts', 'Inst', 'Music', 'Pantry', 'Video']
for domain in domains:
file = open(f'data/{dataset}/{domain}_train.csv', 'r')
u_s = set()
i_s = set()
for line in file.readlines():
ele = line.strip().split(',')
u_s.add(ele[1])
i_s.add(ele[0])
file.close()
for mode in ['val', 'test']:
val_file = open(f'data/{dataset}/{domain}_{mode}.csv', 'r')
val_lines = []
for line in val_file.readlines():
ele = line.strip().split(',')
if ele[1] in u_s and ele[0] in i_s:
val_lines.append(line)
val_file.close()
val_file = open(f'data/{dataset}/{domain}_{mode}.csv', 'w')
for line in val_lines:
val_file.write(line)
val_file.close()
def cal_p_val(array, popmean):
t, p = scipy.stats.ttest_1samp(array, popmean)
return p / 2