-
Notifications
You must be signed in to change notification settings - Fork 12
/
dtcdr.py
211 lines (168 loc) · 11.1 KB
/
dtcdr.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
# -*- coding: utf-8 -*-
# @Time : 2022/3/22
# @Author : Zihan Lin
# @Email : zhlin@ruc.edu.cn
r"""
DTCDR
################################################
Reference:
Feng Zhu et al. "DTCDR: A Framework for Dual-Target Cross-Domain Recommendation." in CIKM 2019.
"""
import numpy as np
import torch
import torch.nn as nn
from recbole.model.init import xavier_normal_initialization
from recbole.utils import InputType
from recbole.model.layers import MLPLayers
from recbole_cdr.model.crossdomain_recommender import CrossDomainRecommender
class DTCDR(CrossDomainRecommender):
r""" This model conduct NeuMF or DMF in both domain where the embedding of overlapped users or items
are combined from both domain.
NOTE:This is the simplified version of original DTCDR model.
To make fair comparison, This implementation only support user ratings in source and target domain.
Other side information (e.g., user comments, user profiles and item details) is not supported.
"""
input_type = InputType.POINTWISE
def __init__(self, config, dataset):
super(DTCDR, self).__init__(config, dataset)
self.SOURCE_LABEL = dataset.source_domain_dataset.label_field
self.TARGET_LABEL = dataset.target_domain_dataset.label_field
# load parameters info
self.embedding_size = config['embedding_size']
self.mlp_hidden_size = config['mlp_hidden_size']
self.dropout_prob = config['dropout_prob']
self.base_model = config['base_model']
self.alpha = config['alpha']
assert self.base_model in ['NeuMF', 'DMF'], "based model {} is not supported! ".format(self.base_model)
# define layers and loss
if self.base_model == 'NeuMF':
self.source_user_embedding = nn.Embedding(self.total_num_users, self.embedding_size)
self.source_item_embedding = nn.Embedding(self.total_num_items, self.embedding_size)
self.target_user_embedding = nn.Embedding(self.total_num_users, self.embedding_size)
self.target_item_embedding = nn.Embedding(self.total_num_items, self.embedding_size)
with torch.no_grad():
self.target_user_embedding.weight[self.target_num_users:].fill_(np.NINF)
self.target_item_embedding.weight[self.target_num_items:].fill_(np.NINF)
self.source_user_embedding.weight[self.overlapped_num_users: self.target_num_users].fill_(np.NINF)
self.source_item_embedding.weight[self.overlapped_num_items: self.target_num_items].fill_(np.NINF)
self.source_mlp_layers = MLPLayers([2 * self.embedding_size] + self.mlp_hidden_size, self.dropout_prob)
self.source_mlp_layers.logger = None # remove logger to use torch.save()
self.source_predict_layer = nn.Linear(self.mlp_hidden_size[-1], 1)
self.target_mlp_layers = MLPLayers([2 * self.embedding_size] + self.mlp_hidden_size, self.dropout_prob)
self.target_mlp_layers.logger = None # remove logger to use torch.save()
self.target_predict_layer = nn.Linear(self.mlp_hidden_size[-1], 1)
else:
self.source_history_user_id, self.source_history_user_value, _ = dataset.history_user_matrix(domain='source')
self.source_history_item_id, self.source_history_item_value, _ = dataset.history_item_matrix(domain='source')
self.source_interaction_matrix = dataset.inter_matrix(form='csr', domain='source').astype(np.float32)
self.source_history_user_id = self.source_history_user_id.to(self.device)
self.source_history_user_value = self.source_history_user_value.to(self.device)
self.source_history_item_id = self.source_history_item_id.to(self.device)
self.source_history_item_value = self.source_history_item_value.to(self.device)
self.target_history_user_id, self.target_history_user_value, _ = dataset.history_user_matrix(domain='target')
self.target_history_item_id, self.target_history_item_value, _ = dataset.history_item_matrix(domain='target')
self.target_interaction_matrix = dataset.inter_matrix(form='csr', domain='target').astype(np.float32)
self.target_history_user_id = self.target_history_user_id.to(self.device)
self.target_history_user_value = self.target_history_user_value.to(self.device)
self.target_history_item_id = self.target_history_item_id.to(self.device)
self.target_history_item_value = self.target_history_item_value.to(self.device)
self.source_user_linear = nn.Linear(in_features=self.source_num_items, out_features=self.embedding_size,
bias=False)
self.source_item_linear = nn.Linear(in_features=self.source_num_users, out_features=self.embedding_size,
bias=False)
self.source_user_fc_layers = MLPLayers([self.embedding_size] + self.mlp_hidden_size)
self.source_item_fc_layers = MLPLayers([self.embedding_size] + self.mlp_hidden_size)
self.target_user_linear = nn.Linear(in_features=self.target_num_items, out_features=self.embedding_size,
bias=False)
self.target_item_linear = nn.Linear(in_features=self.target_num_users, out_features=self.embedding_size,
bias=False)
self.target_user_fc_layers = MLPLayers([self.embedding_size] + self.mlp_hidden_size)
self.target_item_fc_layers = MLPLayers([self.embedding_size] + self.mlp_hidden_size)
self.source_sigmoid = nn.Sigmoid()
self.target_sigmoid = nn.Sigmoid()
self.loss = nn.BCELoss()
self.apply(xavier_normal_initialization)
def forward(self, user, item):
user_e = self.get_user_embedding(user)
item_e = self.get_item_embedding(item)
return self.sigmoid(torch.mul(user_e, item_e).sum(dim=1))
def neumf_forward(self, user, item, domain='source'):
user_source_e = self.source_user_embedding(user)
user_target_e = self.target_user_embedding(user)
user_e = torch.maximum(user_source_e, user_target_e)
item_source_e = self.source_item_embedding(item)
item_target_e = self.target_item_embedding(item)
item_e = torch.maximum(item_source_e, item_target_e)
if domain == 'source':
output = self.source_sigmoid(self.source_predict_layer(self.source_mlp_layers(torch.cat((user_e, item_e), -1))))
else:
output = self.target_sigmoid(self.target_predict_layer(self.target_mlp_layers(torch.cat((user_e, item_e), -1))))
return output.squeeze(-1)
def construct_matrix(self, input_tensor, history_id_matrix, history_value_matrix, length):
col_indices = history_id_matrix[input_tensor].flatten()
row_indices = torch.arange(input_tensor.shape[0]).to(self.device)
row_indices = row_indices.repeat_interleave(history_id_matrix.shape[1], dim=0)
matrix_01 = torch.zeros(1).to(self.device).repeat(input_tensor.shape[0], length)
matrix_01.index_put_((row_indices, col_indices), history_value_matrix[input_tensor].flatten())
return matrix_01
def dmf_forward(self, user, item, domain='source'):
col_indices = self.source_history_item_id[user].flatten()
col_indices[col_indices > self.target_num_items] = col_indices[col_indices > self.target_num_items]-(self.target_num_items-self.overlapped_num_items)
row_indices = torch.arange(user.shape[0]).to(self.device)
row_indices = row_indices.repeat_interleave(self.source_history_item_id.shape[1], dim=0)
source_user_matrix = torch.zeros(1).to(self.device).repeat(user.shape[0], self.source_num_items)
source_user_matrix.index_put_((row_indices, col_indices), self.source_history_item_value[user].flatten())
source_user_e = self.source_user_linear(source_user_matrix)
target_user_matrix = self.construct_matrix(user, self.target_history_item_id, self.target_history_item_value, self.target_num_items)
target_user_e = self.target_user_linear(target_user_matrix)
user_e = torch.maximum(source_user_e, target_user_e)
col_indices = self.source_history_user_id[item].flatten()
col_indices[col_indices > self.target_num_users] = col_indices[col_indices > self.target_num_users] - (
self.target_num_users - self.overlapped_num_users)
row_indices = torch.arange(user.shape[0]).to(self.device)
row_indices = row_indices.repeat_interleave(self.source_history_user_id.shape[1], dim=0)
source_item_matrix = torch.zeros(1).to(self.device).repeat(item.shape[0], self.source_num_users)
source_item_matrix.index_put_((row_indices, col_indices), self.source_history_user_value[user].flatten())
source_item_e = self.source_item_linear(source_item_matrix)
target_item_matrix = self.construct_matrix(item, self.target_history_user_id, self.target_history_user_value, self.target_num_users)
target_item_e = self.target_item_linear(target_item_matrix)
item_e = torch.maximum(source_item_e, target_item_e)
if domain == 'source':
user_e = self.source_user_fc_layers(user_e)
item_e = self.source_item_fc_layers(item_e)
output = torch.mul(user_e, item_e).sum(dim=1)
output = self.source_sigmoid(output)
else:
user_e = self.target_user_fc_layers(user_e)
item_e = self.target_item_fc_layers(item_e)
output = torch.mul(user_e, item_e).sum(dim=1)
output = self.target_sigmoid(output)
return output
def calculate_loss(self, interaction):
source_user = interaction[self.SOURCE_USER_ID]
source_item = interaction[self.SOURCE_ITEM_ID]
source_label = interaction[self.SOURCE_LABEL]
target_user = interaction[self.TARGET_USER_ID]
target_item = interaction[self.TARGET_ITEM_ID]
target_label = interaction[self.TARGET_LABEL]
if self.base_model == 'NeuMF':
source_output = self.neumf_forward(source_user, source_item, 'source')
target_output = self.neumf_forward(target_user, target_item, 'target')
loss_s = self.loss(source_output, source_label)
loss_t = self.loss(target_output, target_label)
return loss_s * self.alpha + loss_t * (1 - self.alpha)
else:
source_output = self.dmf_forward(source_user, source_item, 'source')
target_output = self.dmf_forward(target_user, target_item, 'source')
loss_s = self.loss(source_output, source_label)
loss_t = self.loss(target_output, target_label)
return loss_s * self.alpha + loss_t * (1 - self.alpha)
def predict(self, interaction):
user = interaction[self.TARGET_USER_ID]
item = interaction[self.TARGET_ITEM_ID]
if self.base_model == 'NeuMF':
output = self.neumf_forward(user, item, 'target')
return output
else:
output = self.dmf_forward(user, item, 'target')
return output