-
Notifications
You must be signed in to change notification settings - Fork 13
/
Copy pathsscdr.py
259 lines (228 loc) · 13 KB
/
sscdr.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
# -*- coding: utf-8 -*-
# @Time : 2022/5/13
# @Author : Zihan Lin
# @Email : zhlin@ruc.edu.cn
r"""
SSCDR
################################################
Reference:
SeongKu Kang et al. "Semi-Supervised Learning for Cross-Domain Recommendation to Cold-Start Users" in CIKM 2019.
"""
import numpy as np
import torch
import torch.nn as nn
from recbole.model.init import xavier_normal_initialization
from recbole.model.layers import MLPLayers
from recbole.utils import InputType
from recbole_cdr.model.crossdomain_recommender import CrossDomainRecommender
class SSCDR(CrossDomainRecommender):
r"""SSCDR conducts the embedding mapping by both supervised way and semi-supervised way.
In this implementation, the mapped embedding is used for all the overlapped users (or items) in target domain.
"""
input_type = InputType.PAIRWISE
def __init__(self, config, dataset):
super(SSCDR, self).__init__(config, dataset)
assert self.overlapped_num_items == 1 or self.overlapped_num_users == 1, \
"SSCDR model only support user overlapped or item overlapped dataset! "
if self.overlapped_num_users > 1:
self.mode = 'overlap_users'
elif self.overlapped_num_items > 1:
self.mode = 'overlap_items'
else:
self.mode = 'non_overlap'
self.phase = None
self.dataset = dataset.source_domain_dataset.inter_feat
# load dataset info
self.embedding_size = config['embedding_size']
self.lamda = config['lambda']
self.margin = config['margin']
self.mlp_hidden_size = config['mlp_hidden_size']
self.mapping_layer = MLPLayers(layers=[self.embedding_size] + self.mlp_hidden_size + [self.embedding_size],
activation='tanh', dropout=0, bn=False)
if self.mode == 'overlap_users':
self.user_interacted_items = self.build_interacted_items(dataset, mode='user')
elif self.mode == 'overlap_items':
self.item_interacted_users = self.build_interacted_items(dataset, mode='item')
# define layers and loss
self.source_user_embedding = torch.nn.Embedding(self.total_num_users, self.embedding_size)
self.source_item_embedding = torch.nn.Embedding(self.total_num_items, self.embedding_size)
self.target_user_embedding = torch.nn.Embedding(self.total_num_users, self.embedding_size)
self.target_item_embedding = torch.nn.Embedding(self.total_num_items, self.embedding_size)
with torch.no_grad():
self.source_user_embedding.weight[self.overlapped_num_users: self.target_num_users].fill_(0)
self.source_item_embedding.weight[self.overlapped_num_items: self.target_num_items].fill_(0)
self.target_user_embedding.weight[self.target_num_users:].fill_(0)
self.target_item_embedding.weight[self.target_num_items:].fill_(0)
self.map_loss = nn.MSELoss()
self.rec_loss = nn.TripletMarginLoss(margin=self.margin)
# parameters initialization
self.apply(xavier_normal_initialization)
def build_interacted_items(self, dataset, mode='user'):
dataset = dataset.source_domain_dataset
if mode == 'user':
interacted_items = [[] for _ in range(self.total_num_users)]
for uid, iid in zip(dataset.inter_feat[dataset.uid_field].numpy(),
dataset.inter_feat[dataset.iid_field].numpy()):
interacted_items[uid].append(iid)
return interacted_items
else:
interacted_users = [[] for _ in range(self.total_num_items)]
for iid, uid in zip(dataset.inter_feat[dataset.iid_field].numpy(),
dataset.inter_feat[dataset.uid_field].numpy()):
interacted_users[iid].append(uid)
return interacted_users
def sample(self, ids, mode='user'):
ids = ids.cpu().numpy()
interacted = np.zeros_like(ids)
non_interacted = np.zeros_like(ids)
if mode =='user':
all_candidates = list(range(self.overlapped_num_items)) + \
list(range(self.target_num_items, self.total_num_items))
for index, id in enumerate(ids):
interacted_items = self.user_interacted_items[id]
if len(interacted_items) == 0:
interacted_items.append(0)
non_interacted_id = np.random.choice(all_candidates, size=1)[0]
while non_interacted_id in interacted_items:
non_interacted_id = np.random.choice(all_candidates, size=1)[0]
interacted[index] = np.random.choice(interacted_items, size=1)[0]
non_interacted[index] = non_interacted_id
else:
all_candidates = list(range(self.overlapped_num_users)) + \
list(range(self.target_num_users, self.total_num_users))
for index, id in enumerate(ids):
interacted_users = self.item_interacted_users[id]
if len(interacted_users) == 0:
interacted_users.append(0)
non_interacted_id = np.random.choice(all_candidates, size=1)[0]
while non_interacted_id in interacted_users:
non_interacted_id = np.random.choice(all_candidates, size=1)[0]
interacted[index] = np.random.choice(interacted_users, size=1)[0]
non_interacted[index] = non_interacted_id
return torch.from_numpy(interacted).to(self.device), torch.from_numpy(non_interacted).to(self.device)
@staticmethod
def embedding_normalize(embeddings):
emb_length = torch.sum(embeddings**2, dim=1, keepdim=True)
ones = torch.ones_like(emb_length)
norm = torch.where(emb_length > 1, emb_length, ones)
return embeddings / norm
@staticmethod
def embedding_distance(emb1, emb2):
return torch.sum((emb1-emb2)**2, dim=1)
def set_phase(self, phase):
self.phase = phase
def calculate_source_loss(self, interaction):
source_user = interaction[self.SOURCE_USER_ID]
source_pos_item = interaction[self.SOURCE_ITEM_ID]
source_neg_item = interaction[self.SOURCE_NEG_ITEM_ID]
source_user_e = self.source_user_embedding(source_user)
source_pos_item_e = self.source_item_embedding(source_pos_item)
source_neg_item_e = self.source_item_embedding(source_neg_item)
loss_t = self.rec_loss(self.embedding_normalize(source_user_e),
self.embedding_normalize(source_pos_item_e),
self.embedding_normalize(source_neg_item_e))
return loss_t
def calculate_target_loss(self, interaction):
target_user = interaction[self.TARGET_USER_ID]
target_pos_item = interaction[self.TARGET_ITEM_ID]
target_neg_item = interaction[self.TARGET_NEG_ITEM_ID]
target_user_e = self.target_user_embedding(target_user)
target_pos_item_e = self.target_item_embedding(target_pos_item)
target_neg_item_e = self.target_item_embedding(target_neg_item)
loss_t = self.rec_loss(self.embedding_normalize(target_user_e),
self.embedding_normalize(target_pos_item_e),
self.embedding_normalize(target_neg_item_e))
return loss_t
def calculate_map_loss(self, interaction):
idx = interaction[self.OVERLAP_ID].squeeze(1)
if self.mode == 'overlap_users':
source_user_e = self.source_user_embedding(idx)
target_user_e = self.target_user_embedding(idx)
map_e = self.mapping_layer(source_user_e)
loss_s = self.map_loss(map_e, target_user_e)
source_pos_item, source_neg_item = self.sample(idx, mode='user')
map_pos_item_e = self.mapping_layer(self.source_item_embedding(source_pos_item))
map_neg_item_e = self.mapping_layer(self.source_item_embedding(source_neg_item))
loss_u = self.rec_loss(self.embedding_normalize(target_user_e),
self.embedding_normalize(map_pos_item_e),
self.embedding_normalize(map_neg_item_e))
else:
source_item_e = self.source_item_embedding(idx)
target_item_e = self.target_item_embedding(idx)
map_e = self.mapping_layer(source_item_e)
loss_s = self.map_loss(map_e, target_item_e)
source_pos_user, source_neg_user = self.sample(idx, mode='item')
map_pos_user_e = self.mapping_layer(self.source_user_embedding(source_pos_user))
map_neg_user_e = self.mapping_layer(self.source_user_embedding(source_neg_user))
loss_u = self.rec_loss(self.embedding_normalize(target_item_e),
self.embedding_normalize(map_pos_user_e),
self.embedding_normalize(map_neg_user_e))
return loss_s + self.lamda * loss_u
def calculate_loss(self, interaction):
if self.phase == 'SOURCE':
return self.calculate_source_loss(interaction)
elif self.phase == 'OVERLAP':
return self.calculate_map_loss(interaction)
else:
return self.calculate_target_loss(interaction)
def predict(self, interaction):
if self.phase == 'SOURCE':
user = interaction[self.SOURCE_USER_ID]
item = interaction[self.SOURCE_ITEM_ID]
user_e = self.embedding_normalize(self.source_user_embedding(user))
item_e = self.embedding_normalize(self.source_item_embedding(item))
score = -self.embedding_distance(user_e, item_e)
elif self.phase == 'TARGET':
user = interaction[self.TARGET_USER_ID]
item = interaction[self.TARGET_ITEM_ID]
user_e = self.embedding_normalize(self.target_user_embedding(user))
item_e = self.embedding_normalize(self.target_item_embedding(item))
score = -self.embedding_distance(user_e, item_e)
else:
user = interaction[self.TARGET_USER_ID]
item = interaction[self.TARGET_ITEM_ID]
if self.mode == 'overlap_users':
repeat_user = user.repeat(self.embedding_size, 1).transpose(0, 1)
user_e = torch.where(repeat_user < self.overlapped_num_users, self.mapping_layer(self.source_user_embedding(user)),
self.target_user_embedding(user))
item_e = self.target_item_embedding(item)
else:
user_e = self.target_user_embedding(user)
repeat_item = item.repeat(self.embedding_size, 1).transpose(0, 1)
item_e = torch.where(repeat_item < self.overlapped_num_items, self.mapping_layer(self.source_item_embedding(item)),
self.target_item_embedding(item))
user_e = self.embedding_normalize(user_e)
item_e = self.embedding_normalize(item_e)
score = -self.embedding_distance(user_e, item_e)
return score
def full_sort_predict(self, interaction):
if self.phase == 'SOURCE':
user = interaction[self.SOURCE_USER_ID]
user_e = self.embedding_normalize(self.source_user_embedding(user))
overlap_item_e = self.embedding_normalize(self.source_item_embedding.weight[:self.overlapped_num_items])
source_item_e = self.embedding_normalize(self.source_item_embedding.weight[self.target_num_items:])
all_item_e = torch.cat([overlap_item_e, source_item_e], dim=0)
elif self.phase == 'TARGET':
user = interaction[self.TARGET_USER_ID]
user_e = self.embedding_normalize(self.target_user_embedding(user))
all_item_e = self.embedding_normalize(self.target_item_embedding.weight[:self.target_num_items])
else:
user = interaction[self.TARGET_USER_ID]
if self.mode == 'overlap_users':
repeat_user = user.repeat(self.embedding_size, 1).transpose(0, 1)
user_e = torch.where(repeat_user < self.overlapped_num_users, self.mapping_layer(self.source_user_embedding(user)),
self.target_user_embedding(user))
all_item_e = self.target_item_embedding.weight[:self.target_num_items]
else:
user_e = self.target_user_embedding(user)
overlap_item_e = self.mapping_layer(self.source_item_embedding.weight[:self.overlapped_num_items])
target_item_e = self.target_item_embedding.weight[self.overlapped_num_items:self.target_num_items]
all_item_e = torch.cat([overlap_item_e, target_item_e], dim=0)
user_e = self.embedding_normalize(user_e)
all_item_e = self.embedding_normalize(all_item_e)
num_batch_user, emb_dim = user_e.size()
num_all_item, _ = all_item_e.size()
dist = -2 * torch.matmul(user_e, all_item_e.permute(1, 0))
dist += torch.sum(user_e ** 2, -1).view(num_batch_user, 1)
dist += torch.sum(all_item_e ** 2, -1).view(1, num_all_item)
return -dist.view(-1)