-
Notifications
You must be signed in to change notification settings - Fork 132
/
w2v_content_recall.py
219 lines (173 loc) · 7.65 KB
/
w2v_content_recall.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
import warnings
warnings.filterwarnings('ignore')
import os
import pandas as pd
import numpy as np
from tqdm import tqdm
from gensim.models import Word2Vec
import datetime
from time import time
def gen_detail_content_recall(sim_dict, target_df, data_hist,
uid, iid, time_col,
topn=20, topk=100, prefix='detail'):
def REC(sim_dict, hists, topn=20, topk=100):
rank = {}
for art in hists:
if art not in sim_dict:
continue
cnt = 0
for sart, v in sim_dict[art].items():
rank[sart] = max(rank.get(sart, 0), v)
cnt += 1
if cnt > topn:
break
return sorted(rank.items(), key=lambda d: d[1], reverse=True)[:topk]
df = target_df.copy()
tmp = data_hist.groupby(uid)[iid].agg(list).reset_index()
df = df.merge(tmp, on=uid, how='left')
samples = []
for cur_uid, label, hists in tqdm(df.values):
if hists is np.nan:
continue
rec = REC(sim_dict, hists, topn, topk)
for k, v in rec:
if k in label:
samples.append([cur_uid, k, v, 1])
else:
samples.append([cur_uid, k, v, 0])
samples = pd.DataFrame(samples, columns=[uid, iid, '{}_content_sim_score'.format(prefix), 'label'])
print('{} content recall: '.format(prefix), samples.shape, samples.label.mean())
return samples
def gen_detail_content_recall_test(sim_dict, data_hist,
uid, iid, time_col,
topn=20, topk=100, prefix='detail'):
def REC(sim_dict, hists, topn=20, topk=100):
rank = {}
for art in hists:
if art not in sim_dict:
continue
cnt = 0
for sart, v in sim_dict[art].items():
rank[sart] = max(rank.get(sart, 0), v)
cnt += 1
if cnt > topn:
break
return sorted(rank.items(), key=lambda d: d[1], reverse=True)[:topk]
df = data_hist.groupby(uid)[iid].agg(list).reset_index()
samples = []
for cur_uid, hists in tqdm(df.values):
if hists is np.nan:
continue
rec = REC(sim_dict, hists, topn, topk)
for k, v in rec:
samples.append([cur_uid, k, v])
samples = pd.DataFrame(samples, columns=[uid, iid, '{}_content_sim_score'.format(prefix)])
print('{} content recall: '.format(prefix), samples.shape)
return samples
def get_art_sim_dict(df, art_map_dic,
uid, iid, time_col,
topn=100):
feats = [c for c in df.columns if c not in [iid]]
split_size = 2000
split_num = int(len(df) / split_size)
if len(df) % split_size != 0:
split_num += 1
w2v_vec = df[feats].values
l2norm = np.linalg.norm(w2v_vec, axis=1, keepdims=True)
w2v_vec = w2v_vec / (l2norm + 1e-9)
w2v_vec_T = w2v_vec.T
art_sim_dict = {}
cnt = 0
for i in tqdm(range(split_num)):
vec = w2v_vec[i * split_size:(i + 1) * split_size]
sim = vec.dot(w2v_vec_T)
idx = (-sim).argsort(axis=1)
sim = (-sim)
sim.sort(axis=1)
idx = idx[:, :topn]
score = sim[:, :topn]
score = -score
for idx_, score_ in zip(idx, score):
idx_ = [art_map_dic[j] for j in idx_]
art_sim_dict[art_map_dic[cnt]] = dict(zip(idx_, score_))
cnt += 1
return art_sim_dict
def train_model(data, size=10, save_path='w2v_model/', iter=5, window=20):
"""训练模型"""
print('Begin training w2v model')
begin_time = time()
# if not os.path.exists(save_path):
# os.mkdir(save_path)
# model = Word2Vec(data, vector_size=size, window=window, min_count=0, workers=20,
# seed=1997, epochs=iter, sg=1, hs=1, compute_loss=True)
# print(model.get_latest_training_loss())
model = Word2Vec(sentences=data, vector_size=size, window=window, min_count=1, workers=20)
end_time = time()
run_time = end_time - begin_time
print('该循环程序运行时间:', round(run_time, 2))
return model
def get_w2v_model(df_, date,
uid, iid, time_col,
last_days=30, size=10, iter=5, save_path='w2v_model/', window=20):
begin_date = datetime.datetime.strptime(date, '%Y-%m-%d %H:%M:%S') - datetime.timedelta(days=last_days)
begin_date = str(begin_date)
df = df_[(df_[time_col] <= date) & (df_[time_col] >= begin_date)]
user_item = df.groupby(uid)[iid].agg(list).reset_index()
model = train_model(user_item[iid].values, size=size, iter=iter, save_path=save_path, window=window)
return model
def generate_w2v_sim(transactions_train, date,
uid, iid, time_col,
last_days = 180, size=5):
w2v_model = get_w2v_model(transactions_train, date,
uid, iid, time_col,
size=size, last_days=last_days)
w2v_df = pd.DataFrame()
w2v_df[iid] = w2v_model.wv.index_to_key
w2v_vectors = pd.DataFrame(w2v_model.wv.vectors,
columns=[f'{iid}_w2v_dim{i}' for i in range(w2v_model.wv.vector_size)])
w2v_df = pd.concat([w2v_df, w2v_vectors], axis=1)
pop_num = 6000
begin_date = datetime.datetime.strptime(date, '%Y-%m-%d %H:%M:%S') - datetime.timedelta(days=7)
begin_date = str(begin_date)
data_lw = transactions_train[
(transactions_train[time_col] >= begin_date) & (transactions_train[time_col] <= date)]
dummy_dict = data_lw[iid].value_counts()
recent_active_items = list(dummy_dict.index[:pop_num])
df = w2v_df[w2v_df[iid].isin(recent_active_items)]
art_map_dic = dict(zip(range(len(df)), df[iid].values.tolist()))
art_sim_dict = get_art_sim_dict(df, art_map_dic,
uid, iid, time_col,
topn=200)
return art_sim_dict
def w2v_concent_recall(uids, data, date,
uid, iid, time_col,
last_days=7, dtype='train',
topn=20, topk=100, prefix='w2v'):
assert dtype in ['train', 'test']
if dtype == 'train':
begin_date = datetime.datetime.strptime(date, '%Y-%m-%d %H:%M:%S') - datetime.timedelta(days=last_days)
begin_date = str(begin_date)
target_df = data[(data[time_col] <= date) & (data[time_col] > begin_date)]
print(target_df[time_col].min(), target_df[time_col].max())
target = target_df.groupby(uid)[iid].agg(list).reset_index()
target.columns = [uid, 'label']
data_hist = data[data[time_col] <= begin_date]
data_hist_ = data_hist[data_hist[uid].isin(target[uid].unique())]
last_days, size = 180, 32
sim_dict = generate_w2v_sim(data, date,
uid, iid, time_col,
last_days=last_days, size=size)
samples = gen_detail_content_recall(sim_dict, target, data_hist_,
uid, iid, time_col,
topn=topn, topk=topk, prefix=prefix)
return samples
elif dtype == 'test':
last_days, size = 180, 32
sim_dict = generate_w2v_sim(data, date,
uid, iid, time_col,
last_days=last_days, size=size)
data_ = data[data[uid].isin(uids)]
samples = gen_detail_content_recall_test(sim_dict, data_,
uid, iid, time_col,
topn=topn, topk=topk, prefix=prefix)
return samples