forked from ZJU-DAILY/MetaKG
-
Notifications
You must be signed in to change notification settings - Fork 0
/
MetaKG.py
359 lines (283 loc) · 14.6 KB
/
MetaKG.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
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
import random
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch_scatter import scatter_mean, scatter_sum, scatter_max
from torch_scatter.utils import broadcast
from collections import OrderedDict
class Aggregator(nn.Module):
"""
Relational Path-aware Convolution Network
"""
def __init__(self, n_users, n_items, triplet_attention, use_gate):
super(Aggregator, self).__init__()
self.n_users = n_users
self.n_items = n_items
self.triplet_attention = triplet_attention
self.use_gate = use_gate
self.gate1 = nn.Linear(64, 64, bias=False)
self.gate2 = nn.Linear(64, 64, bias=False)
self.sigmoid = nn.Sigmoid()
def scatter_softmax(self, src, index, dim: int = -1, eps: float = 1e-12):
if not torch.is_floating_point(src):
raise ValueError('`scatter_softmax` can only be computed over tensors '
'with floating point data types.')
index = broadcast(index, src, dim)
max_value_per_index = scatter_max(src, index, dim=dim)[0]
max_per_src_element = max_value_per_index.gather(dim, index)
recentered_scores = src - max_per_src_element
recentered_scores_exp = recentered_scores.exp()
sum_per_index = scatter_sum(recentered_scores_exp, index, dim)
normalizing_constants = sum_per_index.add_(eps).gather(dim, index)
return recentered_scores_exp.div(normalizing_constants)
def KG_forward(self, entity_emb, edge_index, edge_type, weight):
n_entities = entity_emb.shape[0]
head, tail = edge_index
edge_relation_emb = weight[edge_type]
neigh_relation_emb = entity_emb[tail] * edge_relation_emb # [-1, channel]
entity_agg = scatter_mean(src=neigh_relation_emb, index=head, dim_size=n_entities, dim=0)
return entity_agg
def forward(self, entity_emb, user_emb, edge_index,
edge_type, interact_mat, weight, fast_weights=None, i=0):
"""KG aggregate"""
entity_agg = self.KG_forward(entity_emb, edge_index, edge_type, weight)
"""user aggregate"""
if self.use_gate:
item_kg_agg = entity_agg[:self.n_items]
att_kg_agg = entity_agg[self.n_items:]
mat_row = interact_mat._indices()[0, :]
mat_col = interact_mat._indices()[1, :]
mat_val = interact_mat._values()
item_neigh_emb = user_emb[mat_row] * weight[0]
i_u_agg = scatter_mean(src=item_neigh_emb, index=mat_col, dim_size=self.n_items, dim=0)
if fast_weights == None:
gi = self.sigmoid(self.gate1(item_kg_agg) + self.gate2(i_u_agg))
else:
gate1_name = 'convs.{}.gate1.weight'.format(str(i))
gate2_name = 'convs.{}.gate2.weight'.format(str(i))
conv_w1 = fast_weights[gate1_name]
conv_w2 = fast_weights[gate2_name]
gi = self.sigmoid(F.linear(item_kg_agg, conv_w1) + F.linear(i_u_agg, conv_w2))
item_emb_fusion = (gi * item_kg_agg) + ((1 - gi) * i_u_agg)
user_item_mat = torch.sparse.FloatTensor(torch.cat([mat_row, mat_col]).view(2, -1),
torch.ones_like(mat_val),
size=[self.n_users, self.n_items])
user_agg = torch.sparse.mm(user_item_mat, item_emb_fusion)
entity_agg = torch.cat([item_emb_fusion, att_kg_agg])
else:
user_agg = torch.sparse.mm(interact_mat, entity_emb)
return entity_agg, user_agg
class GraphConv(nn.Module):
"""
Graph Convolutional Network
"""
def __init__(self, channel, n_hops, n_users, n_relations, n_items, use_gate,
node_dropout_rate=0.5, mess_dropout_rate=0.1):
super(GraphConv, self).__init__()
self.convs = nn.ModuleList()
self.n_relations = n_relations
self.n_users = n_users
self.n_items = n_items
self.node_dropout_rate = node_dropout_rate
self.mess_dropout_rate = mess_dropout_rate
self.triplet_attention = self.Consis_attention()
weight = nn.init.xavier_uniform_(torch.empty(n_relations, channel)) # not include interact
self.weight = nn.Parameter(weight) # [n_relations - 1, in_channel]
for i in range(n_hops):
self.convs.append(Aggregator(n_users=n_users, n_items=n_items, triplet_attention=self.triplet_attention, use_gate=use_gate))
self.dropout = nn.Dropout(p=mess_dropout_rate) # mess dropout
def Consis_attention(self):
# used in KCAN (CIKM 21), no parameter
return nn.CosineSimilarity(dim=1, eps=1e-6)
def _edge_sampling(self, edge_index, edge_type, rate=0.5):
# edge_index: [2, -1]
# edge_type: [-1]
n_edges = edge_index.shape[1]
random_indices = np.random.choice(n_edges, size=int(n_edges * rate), replace=False)
return edge_index[:, random_indices], edge_type[random_indices]
def _sparse_dropout(self, x, rate=0.5):
noise_shape = x._nnz()
random_tensor = rate
random_tensor += torch.rand(noise_shape).to(x.device)
dropout_mask = torch.floor(random_tensor).type(torch.bool)
i = x._indices()
v = x._values()
i = i[:, dropout_mask]
v = v[dropout_mask]
out = torch.sparse.FloatTensor(i, v, x.shape).to(x.device)
return out * (1. / (1 - rate))
def forward(self, user_emb, entity_emb, edge_index, edge_type,
interact_mat, fast_weights=None, mess_dropout=True, node_dropout=True):
"""node dropout"""
if node_dropout:
edge_index, edge_type = self._edge_sampling(edge_index, edge_type, self.node_dropout_rate)
# interact_mat = self._sparse_dropout(interact_mat, self.node_dropout_rate)
entity_res_emb = entity_emb # [n_entity, channel]
user_res_emb = user_emb # [n_users, channel]
for i in range(len(self.convs)):
entity_emb, user_emb = self.convs[i](entity_emb, user_emb,
edge_index, edge_type, interact_mat,
self.weight, fast_weights, i=i)
"""message dropout"""
if mess_dropout:
entity_emb = self.dropout(entity_emb)
user_emb = self.dropout(user_emb)
entity_emb = F.normalize(entity_emb)
user_emb = F.normalize(user_emb)
"""result emb"""
entity_res_emb = torch.add(entity_res_emb, entity_emb)
user_res_emb = torch.add(user_res_emb, user_emb)
return entity_res_emb, user_res_emb
class Recommender(nn.Module):
def __init__(self, data_config, args_config, graph, user_pre_embed, item_pre_embed):
super(Recommender, self).__init__()
self.n_users = data_config['n_users']
self.n_items = data_config['n_items']
self.n_relations = data_config['n_relations']
self.n_entities = data_config['n_entities']
self.n_nodes = data_config['n_nodes'] # n_users + n_entities
self.user_pre_embed = user_pre_embed
self.item_pre_embed = item_pre_embed
# inner meta-learning update
self.num_inner_update = args_config.num_inner_update
self.meta_update_lr = args_config.meta_update_lr
self.decay = args_config.l2
self.emb_size = args_config.dim
self.context_hops = args_config.context_hops
self.use_gate = args_config.use_gate
self.node_dropout = args_config.node_dropout
self.node_dropout_rate = args_config.node_dropout_rate
self.mess_dropout = args_config.mess_dropout
self.mess_dropout_rate = args_config.mess_dropout_rate
self.device = torch.device("cuda:" + str(args_config.gpu_id)) if args_config.cuda and torch.cuda.is_available() \
else torch.device("cpu")
self.edge_index, self.edge_type = self._get_edges(graph)
self._init_weight()
self.gcn = self._init_model()
self.interact_mat = None
def _init_weight(self):
self.all_embed = nn.init.xavier_uniform_(torch.empty(self.n_nodes, self.emb_size))
if self.user_pre_embed!=None and self.item_pre_embed!=None:
entity_emb = self.all_embed[(self.n_users + self.n_items):,:]
self.all_embed = torch.cat([self.user_pre_embed, self.item_pre_embed, entity_emb])
self.all_embed = nn.Parameter(self.all_embed)
def _init_model(self):
return GraphConv(channel=self.emb_size,
n_hops=self.context_hops,
n_users=self.n_users,
n_relations=self.n_relations,
n_items=self.n_items,
use_gate=self.use_gate,
node_dropout_rate=self.node_dropout_rate,
mess_dropout_rate=self.mess_dropout_rate)
def _get_edges(self, graph):
graph_tensor = torch.tensor(list(graph.edges)) # [-1, 3]
index = graph_tensor[:, :-1] # [-1, 2]
type = graph_tensor[:, -1] # [-1, 1]
return index.t().long().to(self.device), type.long().to(self.device)
def get_parameter(self):
param_dict = dict()
for name, para in self.gcn.named_parameters():
if name.startswith('conv'):
param_dict[name] = para
return OrderedDict(param_dict)
def forward_kg(self, h, r, pos_t, neg_t):
entity_emb = self.all_embed[self.n_users:, :]
h_emb = entity_emb[h]
r_emb = entity_emb[r]
pos_t_emb = entity_emb[pos_t]
neg_t_emb = entity_emb[neg_t]
r_t_pos = pos_t_emb * r_emb
r_t_neg = neg_t_emb * r_emb
pos_score = torch.sum(torch.pow(r_t_pos - h_emb, 2), dim=1)
neg_score = torch.sum(torch.pow(r_t_neg - h_emb, 2), dim=1)
kg_loss = (-1.0) * F.logsigmoid(neg_score - pos_score)
kg_loss = torch.mean(kg_loss)
return kg_loss
def forward_meta(self, support, query, fast_weights=None):
user_s = support[0]
pos_item_s = support[1]
neg_item_s = support[2]
user_q = query[0]
pos_item_q = query[1]
neg_item_q = query[2]
user_emb = self.all_embed[:self.n_users, :]
entity_emb = self.all_embed[self.n_users:, :]
if fast_weights==None:
fast_weights = self.get_parameter()
for i in range(self.num_inner_update):
entity_gcn_emb, user_gcn_emb = self.gcn(user_emb,
entity_emb,
self.edge_index,
self.edge_type,
self.interact_mat,
fast_weights=fast_weights,
mess_dropout=self.mess_dropout,
node_dropout=self.node_dropout)
u_e = user_gcn_emb[user_s]
pos_e, neg_e = entity_gcn_emb[pos_item_s], entity_gcn_emb[neg_item_s]
loss, _, _ = self.create_bpr_loss(u_e, pos_e, neg_e)
gradients = torch.autograd.grad(loss, fast_weights.values(), create_graph=False)
fast_weights = OrderedDict(
(name, param - self.meta_update_lr * grad)
for ((name, param), grad) in zip(fast_weights.items(), gradients)
)
entity_gcn_emb, user_gcn_emb = self.gcn(user_emb,
entity_emb,
self.edge_index,
self.edge_type,
self.interact_mat,
fast_weights=fast_weights,
mess_dropout=self.mess_dropout,
node_dropout=self.node_dropout)
u_e = user_gcn_emb[user_q]
pos_e, neg_e = entity_gcn_emb[pos_item_q], entity_gcn_emb[neg_item_q]
loss, _, _ = self.create_bpr_loss(u_e, pos_e, neg_e)
return loss
def forward(self, batch=None, is_apapt=False):
if is_apapt:
user = batch['users']
pos_item = batch['pos_items']
neg_item = batch['neg_items']
else:
user = batch[0]
pos_item = batch[1]
neg_item = batch[2]
user_emb = self.all_embed[:self.n_users, :]
entity_emb = self.all_embed[self.n_users:, :]
entity_gcn_emb, user_gcn_emb = self.gcn(user_emb,
entity_emb,
self.edge_index,
self.edge_type,
self.interact_mat,
mess_dropout=self.mess_dropout,
node_dropout=self.node_dropout)
u_e = user_gcn_emb[user]
pos_e, neg_e = entity_gcn_emb[pos_item], entity_gcn_emb[neg_item]
loss, _, _ = self.create_bpr_loss(u_e, pos_e, neg_e)
return loss
def generate(self, adapt_fast_weight=None):
user_emb = self.all_embed[:self.n_users, :]
entity_emb = self.all_embed[self.n_users:, :]
entity_gcn_emb, user_gcn_emb = self.gcn(user_emb,
entity_emb,
self.edge_index,
self.edge_type,
self.interact_mat,
fast_weights=adapt_fast_weight,
mess_dropout=False, node_dropout=False)
return entity_gcn_emb, user_gcn_emb
def rating(self, u_g_embeddings, i_g_embeddings):
return torch.matmul(u_g_embeddings, i_g_embeddings.t())
def create_bpr_loss(self, users, pos_items, neg_items):
batch_size = users.shape[0]
pos_scores = torch.sum(torch.mul(users, pos_items), axis=1)
neg_scores = torch.sum(torch.mul(users, neg_items), axis=1)
mf_loss = -1 * torch.mean(nn.LogSigmoid()(pos_scores - neg_scores))
# cul regularizer
regularizer = (torch.norm(users) ** 2
+ torch.norm(pos_items) ** 2
+ torch.norm(neg_items) ** 2) / 2
emb_loss = self.decay * regularizer / batch_size
return mf_loss + emb_loss, mf_loss, emb_loss