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())