def build_model(self):
time_aware_attention = Time_Aware_Attention()
with tf.variable_scope("UserHistoryEncoder"):
user_history = time_aware_attention.self_attention(
enc=self.behavior_list_embedding_dense,
num_units=self.num_units,
num_heads=self.num_heads,
num_blocks=self.num_blocks,
dropout_rate=self.dropout_rate,
is_training=True,
reuse=False,
key_length=self.seq_length,
query_length=self.seq_length,
t_querys=self.time_list,
t_keys=self.time_list,
t_keys_length=self.max_len,
t_querys_length=self.max_len)
long_term_prefernce = gather_indexes(
batch_size=self.now_bacth_data_size,
seq_length=self.max_len,
width=self.num_units,
sequence_tensor=user_history,
positions=self.mask_index)
self.predict_behavior_emb = long_term_prefernce
self.predict_behavior_emb = layer_norm(self.predict_behavior_emb)
self.output()
def build_model(self):
time_aware_attention = Time_Aware_Attention()
self.gru_net_ins = GRU()
with tf.variable_scope("UserHistoryEncoder"):
user_history = time_aware_attention.self_attention(
enc=self.behavior_list_embedding_dense,
num_units=self.num_units,
num_heads=self.num_heads,
num_blocks=self.num_blocks,
dropout_rate=self.dropout_rate,
is_training=True,
reuse=False,
key_length=self.seq_length,
query_length=self.seq_length,
t_querys=self.time_list,
t_keys=self.time_list,
t_keys_length=self.max_len,
t_querys_length=self.max_len)
with tf.variable_scope('ShortTermIntentEncoder'):
self.time_aware_gru_net_input = tf.concat([
self.behavior_list_embedding_dense,
tf.expand_dims(self.timelast_list, 2),
tf.expand_dims(self.timenow_list, 2)
], 2)
self.short_term_intent_temp = self.gru_net_ins.time_aware_gru_net(
hidden_units=self.num_units,
input_data=self.time_aware_gru_net_input,
input_length=tf.add(self.seq_length, -1),
type='new')
self.short_term_intent = gather_indexes(
batch_size=self.now_bacth_data_size,
seq_length=self.max_len,
width=self.num_units,
sequence_tensor=self.short_term_intent_temp,
positions=self.mask_index - 1)
self.short_term_intent = self.short_term_intent
short_term_intent4vallina = tf.expand_dims(self.short_term_intent,
1)
with tf.variable_scope('NextItemDecoder'):
hybird_preference = time_aware_attention.vanilla_attention(
user_history,
short_term_intent4vallina,
self.num_units,
self.num_heads,
self.num_blocks,
self.dropout_rate,
is_training=True,
reuse=False,
key_length=self.seq_length,
query_length=tf.ones_like(short_term_intent4vallina[:, 0, 0],
dtype=tf.int32),
t_querys=tf.expand_dims(self.target[2], 1),
t_keys=self.time_list,
t_keys_length=self.max_len,
t_querys_length=1)
self.predict_behavior_emb = layer_norm(hybird_preference)
self.output()
def build_model(self):
time_aware_attention = Time_Aware_Attention()
with tf.variable_scope("UserHistoryEncoder"):
user_history = time_aware_attention.self_attention(
enc=self.behavior_list_embedding_dense,
num_units=128,
num_heads=self.num_heads,
num_blocks=self.num_blocks,
dropout_rate=self.dropout_rate,
is_training=True,
reuse=False,
key_length=self.seq_length,
query_length=self.seq_length,
t_querys=self.time_list,
t_keys=self.time_list,
t_keys_length=self.max_len,
t_querys_length=self.max_len)
long_term_prefernce = gather_indexes(
batch_size=self.now_bacth_data_size,
seq_length=self.max_len,
width=self.FLAGS.num_units,
sequence_tensor=user_history,
positions=self.mask_index)
self.predict_behavior_emb = long_term_prefernce
self.predict_behavior_emb = layer_norm(self.predict_behavior_emb)
with tf.name_scope('CrossEntropyLoss'):
l2_norm = tf.add_n([
tf.nn.l2_loss(self.item_list_emb),
tf.nn.l2_loss(self.category_list_emb),
tf.nn.l2_loss(self.position_list_emb)
])
regulation_rate = self.FLAGS.regulation_rate
item_lookup_table_T = tf.transpose(
self.embedding.item_emb_lookup_table)
logits = tf.matmul(self.predict_behavior_emb, item_lookup_table_T)
log_probs = tf.nn.log_softmax(logits)
label_ids = tf.reshape(self.target[0], [-1])
one_hot_labels = tf.one_hot(label_ids,
depth=self.embedding.item_count + 3,
dtype=tf.float32)
self.loss_origin = -tf.reduce_sum(log_probs * one_hot_labels,
axis=[-1])
self.loss = regulation_rate * l2_norm + tf.reduce_mean(
self.loss_origin)
tf.summary.scalar('l2_norm', l2_norm)
tf.summary.scalar('Training Loss', self.loss)
tf.summary.scalar('Learning_rate', self.learning_rate)
self.cal_gradient(tf.trainable_variables())
def build_model(self):
self.gru_net_ins = GRU()
self.ggnn_model = gated_GNN()
with tf.variable_scope('user_behavior_emb'):
user_behavior_list_embedding = self.behavior_list_embedding_dense
with tf.variable_scope('ggnn_encoding'):
gnn_emb = self.ggnn_model.generate_graph_emb(
init_emb=user_behavior_list_embedding,
now_batch_size=self.now_bacth_data_size,
num_units=self.num_units,
adj_in=self.adj_in,
adj_out=self.adj_out,
step=1)
time_aware_attention = Time_Aware_Attention()
with tf.variable_scope("UserHistoryEncoder"):
user_history = time_aware_attention.self_attention(
gnn_emb,
self.num_units,
self.num_heads,
self.num_blocks,
self.dropout_rate,
is_training=True,
reuse=False,
key_length=self.seq_length,
query_length=self.seq_length,
t_querys=self.time_list,
t_keys=self.time_list,
t_keys_length=self.max_len,
t_querys_length=self.max_len)
long_term_prefernce = gather_indexes(
batch_size=self.now_bacth_data_size,
seq_length=self.max_len,
width=self.FLAGS.num_units,
sequence_tensor=user_history,
positions=self.mask_index)
self.predict_behavior_emb = long_term_prefernce
self.predict_behavior_emb = layer_norm(self.predict_behavior_emb)
self.output()
def build_model(self):
time_aware_attention = Time_Aware_Attention()
self.gru_net_ins = GRU()
with tf.variable_scope("UserHistoryEncoder"):
user_history = time_aware_attention.self_attention(
enc=self.behavior_list_embedding_dense,
num_units=self.num_units,
num_heads=self.num_heads,
num_blocks=self.num_blocks,
dropout_rate=self.dropout_rate,
is_training=True,
reuse=False,
key_length=self.seq_length,
query_length=self.seq_length,
t_querys=self.time_list,
t_keys=self.time_list,
t_keys_length=self.max_len,
t_querys_length=self.max_len)
with tf.variable_scope('ShortTermIntentEncoder'):
self.time_aware_gru_net_input = tf.concat([
self.behavior_list_embedding_dense,
tf.expand_dims(self.timelast_list, 2),
tf.expand_dims(self.timenow_list, 2)
], 2)
self.short_term_intent_temp = self.gru_net_ins.time_aware_gru_net(
hidden_units=self.num_units,
input_data=self.time_aware_gru_net_input,
input_length=tf.add(self.seq_length, -1),
type='time-aware')
self.short_term_intent = gather_indexes(
batch_size=self.now_bacth_data_size,
seq_length=self.max_len,
width=self.num_units,
sequence_tensor=self.short_term_intent_temp,
positions=self.mask_index - 1)
self.short_term_intent = self.short_term_intent
short_term_intent4vallina = tf.expand_dims(self.short_term_intent,
1)
with tf.variable_scope('NextItemDecoder'):
hybird_preference = time_aware_attention.vanilla_attention(
user_history,
short_term_intent4vallina,
self.num_units,
self.num_heads,
self.num_blocks,
self.dropout_rate,
is_training=True,
reuse=False,
key_length=self.seq_length,
query_length=tf.ones_like(short_term_intent4vallina[:, 0, 0],
dtype=tf.int32),
t_querys=tf.expand_dims(self.target[2], 1),
t_keys=self.time_list,
t_keys_length=self.max_len,
t_querys_length=1)
self.predict_behavior_emb = hybird_preference
self.predict_behavior_emb = layer_norm(self.predict_behavior_emb)
with tf.variable_scope('OutputLayer'):
long_term_prefernce = gather_indexes(
batch_size=self.now_bacth_data_size,
seq_length=self.max_len,
width=self.num_units,
sequence_tensor=user_history,
positions=self.mask_index)
self.long_term_prefernce = long_term_prefernce
self.short_term_intent = self.short_term_intent
self.hybird_preference = hybird_preference
self.z_concate = tf.concat([
self.long_term_prefernce, self.short_term_intent,
self.hybird_preference
], 1)
self.z = tf.layers.dense(
inputs=self.z_concate,
units=3,
kernel_regularizer=tf.contrib.layers.l2_regularizer(
self.regulation_rate))
self.z = tf.nn.softmax(self.z)
if self.FLAGS.PISTRec_type == 'hard':
if tf.argmax(self.z) == 0:
self.predict_behavior_emb = self.long_term_prefernce
elif tf.argmax(self.z) == 1:
self.predict_behavior_emb = self.short_term_intent
else:
self.predict_behavior_emb = self.hybird_preference
elif self.FLAGS.PISTRec_type == 'soft':
self.predict_behavior_emb = tf.multiply(tf.slice(self.z,[0,0],[-1,1]),self.long_term_prefernce)+\
tf.multiply(tf.slice(self.z, [0, 1], [-1, 1]), self.short_term_intent)+\
tf.multiply(tf.slice(self.z, [0, 2], [-1, 1]), self.hybird_preference)
elif self.FLAGS.PISTRec_type == 'short':
self.predict_behavior_emb = self.short_term_intent
elif self.FLAGS.PISTRec_type == 'long':
self.predict_behavior_emb = self.long_term_prefernce
elif self.FLAGS.PISTRec_type == 'hybird':
self.predict_behavior_emb = self.hybird_preference
self.predict_behavior_emb = layer_norm(self.predict_behavior_emb)
self.output()
def build_model(self):
num_units = self.FLAGS.num_units
num_heads = self.FLAGS.num_heads
num_blocks = self.FLAGS.num_blocks
dropout_rate = self.FLAGS.dropout
time_aware_attention = Time_Aware_Attention()
self.gru_net_ins = GRU()
self.user_embedding, \
self.behavior_list_embedding_dense, \
self.item_list_emb,\
self.category_list_emb, \
self.position_list_emb,\
self.time_list, \
self.timelast_list, \
self.timenow_list, \
self.target, \
self.seq_length = self.embedding.get_embedding(num_units)
max_len = self.FLAGS.length_of_user_history
self.mask_index = tf.reshape(self.seq_length - 1,
[self.now_bacth_data_size, 1])
with tf.variable_scope("UserHistoryEncoder"):
user_history = time_aware_attention.self_attention(
enc=self.behavior_list_embedding_dense,
num_units=128,
num_heads=num_heads,
num_blocks=num_blocks,
dropout_rate=dropout_rate,
is_training=True,
reuse=False,
key_length=self.seq_length,
query_length=self.seq_length,
t_querys=self.time_list,
t_keys=self.time_list,
t_keys_length=max_len,
t_querys_length=max_len)
self.user_h = user_history
with tf.variable_scope('ShortTermIntentEncoder'):
timelast_list = tf.expand_dims(self.timelast_list, 2)
self.time_aware_gru_net_input = tf.concat([
self.behavior_list_embedding_dense,
tf.expand_dims(self.timelast_list, 2),
tf.expand_dims(self.timenow_list, 2)
], 2)
self.error = tf.reshape(tf.add(self.seq_length, -1), [-1])
self.short_term_intent_temp = self.gru_net_ins.time_aware_gru_net(
hidden_units=num_units,
input_data=self.time_aware_gru_net_input,
input_length=tf.add(self.seq_length, -1))
self.short_term_intent = gather_indexes(
batch_size=self.now_bacth_data_size,
seq_length=self.FLAGS.max_len,
width=self.FLAGS.num_units,
sequence_tensor=self.short_term_intent_temp,
positions=self.mask_index - 1)
short_term_intent4vallina = tf.expand_dims(self.short_term_intent,
1)
with tf.variable_scope('NextItemDecoder'):
hybird_preference = time_aware_attention.vanilla_attention(
user_history,
short_term_intent4vallina,
num_units,
num_heads,
num_blocks,
dropout_rate,
is_training=True,
reuse=False,
key_length=self.seq_length,
query_length=tf.ones_like(short_term_intent4vallina[:, 0, 0],
dtype=tf.int32),
t_querys=tf.expand_dims(self.target[2], 1),
t_keys=self.time_list,
t_keys_length=max_len,
t_querys_length=1)
#self.hybird_preference = hybird_preference
with tf.variable_scope('OutputLayer'):
long_term_prefernce = gather_indexes(
batch_size=self.now_bacth_data_size,
seq_length=self.FLAGS.max_len,
width=self.FLAGS.num_units,
sequence_tensor=user_history,
positions=self.mask_index)
self.long_term_prefernce = long_term_prefernce
self.short_term_intent = self.short_term_intent
self.hybird_preference = hybird_preference
self.z_concate = tf.concat([
self.long_term_prefernce, self.short_term_intent,
self.hybird_preference
], 1)
self.z = tf.layers.dense(
inputs=self.z_concate,
units=3,
kernel_regularizer=tf.contrib.layers.l2_regularizer(
self.FLAGS.regulation_rate))
self.z = tf.nn.softmax(self.z)
if self.FLAGS.PISTRec_type == 'hard':
if tf.argmax(self.z) == 0:
self.predict_behavior_emb = self.long_term_prefernce
elif tf.argmax(self.z) == 1:
self.predict_behavior_emb = self.short_term_intent
else:
self.predict_behavior_emb = self.hybird_preference
elif self.FLAGS.PISTRec_type == 'soft':
self.predict_behavior_emb = tf.multiply(tf.slice(self.z,[0,0],[-1,1]),self.long_term_prefernce)+\
tf.multiply(tf.slice(self.z, [0, 1], [-1, 1]), self.short_term_intent)+\
tf.multiply(tf.slice(self.z, [0, 2], [-1, 1]), self.hybird_preference)
elif self.FLAGS.PISTRec_type == 'short':
self.predict_behavior_emb = self.short_term_intent
elif self.FLAGS.PISTRec_type == 'long':
self.predict_behavior_emb = self.long_term_prefernce
elif self.FLAGS.PISTRec_type == 'hybird':
self.predict_behavior_emb = self.hybird_preference
self.predict_behavior_emb = layer_norm(self.predict_behavior_emb)
l2_norm = tf.add_n([
tf.nn.l2_loss(self.item_list_emb),
tf.nn.l2_loss(self.category_list_emb),
tf.nn.l2_loss(self.position_list_emb)
])
regulation_rate = self.FLAGS.regulation_rate
item_lookup_table_T = tf.transpose(
self.embedding.item_emb_lookup_table)
logits = tf.matmul(self.predict_behavior_emb, item_lookup_table_T)
log_probs = tf.nn.log_softmax(logits)
label_ids = tf.reshape(self.target[0], [-1])
one_hot_labels = tf.one_hot(label_ids,
depth=self.embedding.item_count + 3,
dtype=tf.float32)
self.loss_origin = -tf.reduce_sum(log_probs * one_hot_labels,
axis=[-1])
pistloss = regulation_rate * l2_norm + tf.reduce_mean(
self.loss_origin)
with tf.name_scope('LearningtoRankLoss'):
self.loss = pistloss
tf.summary.scalar('l2_norm', l2_norm)
tf.summary.scalar('Training Loss', self.loss)
tf.summary.scalar('Learning_rate', self.learning_rate)
self.cal_gradient(tf.trainable_variables())