-
Notifications
You must be signed in to change notification settings - Fork 4
/
GES_SASRec.py
186 lines (150 loc) · 9.24 KB
/
GES_SASRec.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
'''
GES_SASRec
@author: Tianyu Zhu
@created: 2021/6/19
'''
import tensorflow as tf
class GES_SASRec(object):
def __init__(self, adj_matrix, num_user, num_item, args):
print('model preparing...')
self.adj_matrix = tf.SparseTensor(adj_matrix[0], adj_matrix[1], adj_matrix[2])
self.num_user = num_user
self.num_item = num_item
self.num_factor = args.num_factor
self.l2_reg = args.l2_reg
self.lr = args.lr
self.max_len = args.max_len
self.num_block = args.num_block
self.num_head = args.num_head
self.num_layer = args.num_layer
self.emb_dropout_rate = tf.placeholder(tf.float32)
self.node_dropout_rate = tf.placeholder(tf.float32)
self.input_seq = tf.placeholder(tf.int32, [None, self.max_len], name='input_seq')
self.pos_seq = tf.placeholder(tf.int32, [None, self.max_len], name='pos_seq')
self.neg_seq = tf.placeholder(tf.int32, [None, self.max_len], name='neg_seq')
mask = tf.expand_dims(tf.to_float(tf.not_equal(self.input_seq, self.num_item)), -1)
with tf.name_scope('item_embedding'):
item_embedding = tf.Variable(tf.random_normal([self.num_item, self.num_factor], stddev=0.01, name='item_embedding'))
with tf.name_scope('graph_convolution'):
adj_matrix_dropout = self.node_dropout(self.adj_matrix, len(adj_matrix[0]), 1-self.node_dropout_rate)
item_embedding_final = [item_embedding]
layer = item_embedding
if args.gnn == 'gcn':
W = list()
for k in range(self.num_layer):
W.append(tf.Variable(tf.random_normal([self.num_factor, self.num_factor], stddev=0.01), name='W'+str(k)))
layer = tf.nn.tanh(tf.matmul(tf.sparse_tensor_dense_matmul(adj_matrix_dropout, layer), W[k]))
item_embedding_final += [layer]
elif args.gnn == 'sgc':
for k in range(self.num_layer):
layer = tf.sparse_tensor_dense_matmul(adj_matrix_dropout, layer)
item_embedding_final += [layer]
else:
pass
with tf.name_scope('layer_aggregation'):
if args.layer_agg == 'sum':
item_embedding_final = tf.reduce_sum(tf.stack(item_embedding_final, 1), 1)
elif args.layer_agg == 'avg':
item_embedding_final = tf.reduce_mean(tf.stack(item_embedding_final, 1), 1)
elif args.layer_agg == 'concat':
item_embedding_final = tf.concat(item_embedding_final, 1)
self.num_factor *= (self.num_layer + 1)
else:
item_embedding_final = item_embedding_final[-1]
item_embedding_final_ = tf.concat([item_embedding_final, tf.zeros([1, self.num_factor])], 0)
with tf.name_scope('positional_embedding'):
position = tf.tile(tf.expand_dims(tf.range(self.max_len), 0), [tf.shape(self.input_seq)[0], 1])
position_embedding = tf.Variable(tf.random_normal([self.max_len, self.num_factor], stddev=0.01), name='position_embedding')
p_emb = tf.nn.embedding_lookup(position_embedding, position)
with tf.name_scope('dropout'):
self.seq = tf.nn.embedding_lookup(item_embedding_final_, self.input_seq) * (self.num_factor ** 0.5) + p_emb
self.seq = tf.nn.dropout(self.seq, keep_prob=1-self.emb_dropout_rate) * mask
with tf.name_scope('block'):
for _ in range(self.num_block):
# Self-attention
# Linear projections
seq = self.seq
seq_norm = self.layer_normalize(seq)
Q = tf.layers.dense(seq_norm, self.num_factor, activation=None)
K = tf.layers.dense(seq, self.num_factor, activation=None)
V = tf.layers.dense(seq, self.num_factor, activation=None)
# Split and concat
Q_ = tf.concat(tf.split(Q, self.num_head, axis=2), axis=0)
K_ = tf.concat(tf.split(K, self.num_head, axis=2), axis=0)
V_ = tf.concat(tf.split(V, self.num_head, axis=2), axis=0)
# Multiplication
outputs = tf.matmul(Q_, tf.transpose(K_, [0, 2, 1]))
# Scale
outputs = outputs / (K_.get_shape().as_list()[-1] ** 0.5)
# Key Masking
key_masks = tf.sign(tf.reduce_sum(tf.abs(seq), axis=-1))
key_masks = tf.tile(key_masks, [self.num_head, 1])
key_masks = tf.tile(tf.expand_dims(key_masks, 1), [1, tf.shape(seq_norm)[1], 1])
paddings = tf.ones_like(outputs)*(-2**32+1)
outputs = tf.where(tf.equal(key_masks, 0), paddings, outputs)
# Causality (Future blinding)
diag_vals = tf.ones_like(outputs[0, :, :])
tril = tf.linalg.LinearOperatorLowerTriangular(diag_vals).to_dense()
# tril = tf.contrib.linalg.LinearOperatorTriL(diag_vals).to_dense() # for earlier tf versions
masks = tf.tile(tf.expand_dims(tril, 0), [tf.shape(outputs)[0], 1, 1])
paddings = tf.ones_like(masks)*(-2**32+1)
outputs = tf.where(tf.equal(masks, 0), paddings, outputs)
# Activation
outputs = tf.nn.softmax(outputs)
# Query Masking
query_masks = tf.sign(tf.reduce_sum(tf.abs(seq_norm), axis=-1))
query_masks = tf.tile(query_masks, [self.num_head, 1])
query_masks = tf.tile(tf.expand_dims(query_masks, -1), [1, 1, tf.shape(seq)[1]])
outputs *= query_masks
# Dropouts
outputs = tf.nn.dropout(outputs, keep_prob=1-self.emb_dropout_rate)
# Weighted sum
outputs = tf.matmul(outputs, V_)
# Restore shape
outputs = tf.concat(tf.split(outputs, self.num_head, axis=0), axis=2)
# Residual connection
outputs += seq_norm
# Layer normalization
outputs = self.layer_normalize(outputs)
# Feed forward
# Layer 1
outputs_ = tf.layers.dense(outputs, self.num_factor, activation=tf.nn.relu, use_bias=True)
outputs_ = tf.nn.dropout(outputs_, keep_prob=1-self.emb_dropout_rate)
# Layer 2
outputs_ = tf.layers.dense(outputs_, self.num_factor, activation=None, use_bias=True)
outputs_ = tf.nn.dropout(outputs_, keep_prob=1-self.emb_dropout_rate)
# Residual connection
outputs_ += outputs
self.seq = outputs_ * mask
self.seq = self.layer_normalize(self.seq)
with tf.name_scope('train'):
input_seq_emb = tf.reshape(self.seq, [tf.shape(self.input_seq)[0] * self.max_len, self.num_factor])
pos_seq_emb = tf.nn.embedding_lookup(item_embedding_final_, tf.reshape(self.pos_seq, [tf.shape(self.input_seq)[0] * self.max_len]))
neg_seq_emb = tf.nn.embedding_lookup(item_embedding_final_, tf.reshape(self.neg_seq, [tf.shape(self.input_seq)[0] * self.max_len]))
pos_logits = tf.reduce_sum(pos_seq_emb * input_seq_emb, -1)
neg_logits = tf.reduce_sum(neg_seq_emb * input_seq_emb, -1)
target = tf.reshape(tf.to_float(tf.not_equal(self.pos_seq, self.num_item)), [tf.shape(self.input_seq)[0] * self.max_len])
loss = -tf.reduce_sum(tf.log(tf.sigmoid(pos_logits) + 1e-24) * target + tf.log(1 - tf.sigmoid(neg_logits) + 1e-24) * target) / tf.reduce_sum(target)
self.loss = loss + self.l2_reg * tf.reduce_sum([tf.nn.l2_loss(va) for va in tf.trainable_variables()])
self.train_op = tf.train.AdamOptimizer(self.lr).minimize(self.loss)
with tf.name_scope('test'):
self.test_item = tf.placeholder(tf.int32, [None, 1000])
test_item_emb = tf.nn.embedding_lookup(item_embedding_final, self.test_item) # [batch_size, 1000, self.num_factor]
input_seq_emb_reshape = tf.reshape(input_seq_emb, [tf.shape(self.input_seq)[0], self.max_len, self.num_factor])
input_seq_emb_last = tf.expand_dims(input_seq_emb_reshape[:, -1, :], 1) # [batch_size, 1, self.num_factor]
test_logits = tf.matmul(input_seq_emb_last, tf.transpose(test_item_emb, [0, 2, 1]))
self.test_logits = tf.squeeze(test_logits, 1)
def layer_normalize(self, inputs):
mean, variance = tf.nn.moments(inputs, [-1], keep_dims=True)
beta = tf.Variable(tf.zeros(self.num_factor))
gamma = tf.Variable(tf.ones(self.num_factor))
normalized = (inputs - mean) / ((variance + 1e-24) ** 0.5)
outputs = gamma * normalized + beta
return outputs
def node_dropout(self, adj_matrix, num_value, keep_prob):
noise_shape = [num_value]
random_tensor = keep_prob
random_tensor += tf.random_uniform(noise_shape)
dropout_mask = tf.cast(tf.floor(random_tensor), dtype=tf.bool)
pre_out = tf.sparse_retain(adj_matrix, dropout_mask) * tf.div(1.0, keep_prob)
return pre_out