forked from RUCAIBox/RecBole-CDR
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathnatr.py
191 lines (168 loc) · 9.34 KB
/
natr.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
# -*- coding: utf-8 -*-
# @Time : 2022/5/21
# @Author : Shanlei Mu
# @Email : [email protected]
r"""
NATR
################################################
Reference:
Chen Gao et al. "Cross-domain Recommendation Without Sharing User-relevant Data" in WWW 2019.
"""
import torch
import torch.nn as nn
from recbole.utils import InputType
from recbole.model.loss import RegLoss
from recbole.model.init import xavier_normal_initialization
from recbole_cdr.model.crossdomain_recommender import CrossDomainRecommender
class NATR(CrossDomainRecommender):
r"""NATR propose a neural network method, combining item-level and domain-level
attention mechanisms to address the challenges in cross-domain learning.
"""
input_type = InputType.POINTWISE
def __init__(self, config, dataset):
super(NATR, self).__init__(config, dataset)
assert self.overlapped_num_items == 1 or self.overlapped_num_users == 1, \
"NATR 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
# load parameters info
self.source_embedding_size = config['source_embedding_size']
self.target_embedding_size = config['target_embedding_size']
self.reg_weight = config['reg_weight']
self.max_inter_length = config['max_inter_length']
# load dataset info
self.SOURCE_LABEL = dataset.source_domain_dataset.label_field
self.TARGET_LABEL = dataset.target_domain_dataset.label_field
if self.mode == 'overlap_users':
self.history_user_matrix, self.history_lens, self.mask_mat = self.get_history_user_info(dataset)
if self.mode == 'overlap_items':
self.history_item_matrix, self.history_lens, self.mask_mat = self.get_history_item_info(dataset)
# define layers and loss
self.source_user_embedding = nn.Embedding(self.total_num_users, self.source_embedding_size)
self.source_item_embedding = nn.Embedding(self.total_num_items, self.source_embedding_size)
self.target_user_embedding = nn.Embedding(self.total_num_users, self.target_embedding_size)
self.target_item_embedding = nn.Embedding(self.total_num_items, self.target_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.transfer_layer = nn.Linear(self.source_embedding_size, self.target_embedding_size)
self.unit_attention_layer = nn.Linear(self.target_embedding_size, 1)
self.domain_attention_layer = nn.Linear(self.target_embedding_size, 1)
self.sigmoid = nn.Sigmoid()
self.relu = nn.ReLU()
self.softmax = nn.Softmax(dim=1)
self.rec_loss = nn.BCELoss()
self.reg_loss = RegLoss()
# parameters initialization
self.apply(xavier_normal_initialization)
def set_phase(self, phase):
self.phase = phase
if phase == 'TARGET':
self.source_item_embedding.weight.requires_grad = False
self.source_user_embedding.weight.requires_grad = False
def get_history_item_info(self, dataset):
history_item_matrix, _, history_lens = dataset.history_item_matrix(domain='target')
history_item_matrix = history_item_matrix[:, :self.max_inter_length]
history_item_matrix = history_item_matrix.to(self.device)
history_lens = history_lens.to(self.device)
arange_tensor = torch.arange(history_item_matrix.shape[1]).to(self.device)
mask_mat = (arange_tensor < history_lens.unsqueeze(1)).float()
return history_item_matrix, history_lens, mask_mat
def get_history_user_info(self, dataset):
history_user_matrix, _, history_lens = dataset.history_user_matrix(domain='target')
history_user_matrix = history_user_matrix[:, :self.max_inter_length]
history_user_matrix = history_user_matrix.to(self.device)
history_lens = history_lens.to(self.device)
arange_tensor = torch.arange(history_user_matrix.shape[1]).to(self.device)
mask_mat = (arange_tensor < history_lens.unsqueeze(1)).float()
return history_user_matrix, history_lens, mask_mat
def phase1_forward(self, user, item):
user_e = self.source_user_embedding(user)
item_e = self.source_item_embedding(item)
score = self.sigmoid(torch.mul(user_e, item_e).sum(dim=1))
return score
def calculate_phase1_loss(self, interaction):
user = interaction[self.SOURCE_USER_ID]
item = interaction[self.SOURCE_ITEM_ID]
label = interaction[self.SOURCE_LABEL]
score = self.phase1_forward(user, item)
loss = self.rec_loss(score, label)
return loss
def phase2_forward(self, user, item):
user_e = self.target_user_embedding(user) # (batch, embedding_size)
item_e = self.target_item_embedding(item) # (batch, embedding_size)
if self.mode == 'overlap_items':
batch_mask_mat = self.mask_mat[user]
batch_mask_mat = torch.where(batch_mask_mat.bool(), 0., -10000.0)
history_items = self.history_item_matrix[user]
history_items_e = self.source_item_embedding(history_items) # (batch, n_history, source_embedding_size)
history_items_e = self.transfer_layer(history_items_e)
unit_attention_score = user_e.unsqueeze(1).expand_as(history_items_e) * history_items_e # (batch, n_history, embedding_size)
unit_attention_score = self.unit_attention_layer(self.relu(unit_attention_score)).squeeze(2)
unit_attention_score += batch_mask_mat
unit_attention_score = self.softmax(unit_attention_score) # (batch, n_history)
unit_attention_score = unit_attention_score.unsqueeze(1) # (batch, 1, n_history)
su = torch.bmm(unit_attention_score, history_items_e).squeeze(1) # (batch, embedding_size)
pu, qi = user_e, item_e
b_s = self.domain_attention_layer(self.relu(su * qi))
b_p = self.domain_attention_layer(self.relu(pu * qi))
beta_s = torch.exp(b_s) / (torch.exp(b_s) + torch.exp(b_p))
beta_p = 1 - beta_s
zu = beta_s * su + beta_p * pu # (batch, embedding_size)
score = self.sigmoid(torch.mul(zu, qi).sum(dim=1))
return score
elif self.mode == 'overlap_users':
batch_mask_mat = self.mask_mat[item]
batch_mask_mat = torch.where(batch_mask_mat.bool(), 0., -10000.0)
history_users = self.history_user_matrix[item]
history_users_e = self.source_user_embedding(history_users)
history_users_e = self.transfer_layer(history_users_e)
unit_attention_score = item_e.unsqueeze(1).expand_as(history_users_e) * history_users_e # (batch, n_history, embedding_size)
unit_attention_score = self.unit_attention_layer(self.relu(unit_attention_score)).squeeze(2)
unit_attention_score += batch_mask_mat
unit_attention_score = self.softmax(unit_attention_score) # (batch, n_history)
unit_attention_score = unit_attention_score.unsqueeze(1) # (batch, 1, n_history)
su = torch.bmm(unit_attention_score, history_users_e).squeeze(1) # (batch, embedding_size)
pu, qi = item_e, user_e
b_s = self.domain_attention_layer(self.relu(su * qi))
b_p = self.domain_attention_layer(self.relu(pu * qi))
beta_s = torch.exp(b_s) / (torch.exp(b_s) + torch.exp(b_p))
beta_p = 1 - beta_s
zu = beta_s * su + beta_p * pu # (batch, embedding_size)
score = self.sigmoid(torch.mul(zu, qi).sum(dim=1))
return score
def calculate_phase2_loss(self, interaction):
user = interaction[self.TARGET_USER_ID]
item = interaction[self.TARGET_ITEM_ID]
label = interaction[self.TARGET_LABEL]
score = self.phase2_forward(user, item)
rec_loss = self.rec_loss(score, label)
reg_loss = self.reg_loss([self.target_user_embedding.weight, self.target_item_embedding.weight,
self.transfer_layer.weight,
self.unit_attention_layer.weight, self.domain_attention_layer.weight])
loss = rec_loss + self.reg_weight * reg_loss
return loss
def calculate_loss(self, interaction):
if self.phase == 'SOURCE':
return self.calculate_phase1_loss(interaction)
elif self.phase == 'TARGET':
return self.calculate_phase2_loss(interaction)
else:
return None
def predict(self, interaction):
if self.phase == 'SOURCE':
user = interaction[self.SOURCE_USER_ID]
item = interaction[self.SOURCE_ITEM_ID]
score = self.phase1_forward(user, item)
else:
user = interaction[self.TARGET_USER_ID]
item = interaction[self.TARGET_ITEM_ID]
score = self.phase2_forward(user, item)
return score