def _build_model(self):
self.user_indices = tf.placeholder(tf.int32, shape=[None])
self.item_indices = tf.placeholder(tf.int32, shape=[None])
self.labels = tf.placeholder(tf.float32, shape=[None])
self.bu_var = tf.get_variable(name="bu_var",
shape=[self.n_users],
initializer=tf_zeros,
regularizer=self.reg)
self.bi_var = tf.get_variable(name="bi_var",
shape=[self.n_items],
initializer=tf_zeros,
regularizer=self.reg)
self.pu_var = tf.get_variable(name="pu_var",
shape=[self.n_users, self.embed_size],
initializer=tf_truncated_normal(
0.0, 0.05),
regularizer=self.reg)
self.qi_var = tf.get_variable(name="pi_var",
shape=[self.n_items, self.embed_size],
initializer=tf_truncated_normal(
0.0, 0.05),
regularizer=self.reg)
bias_user = tf.nn.embedding_lookup(self.bu_var, self.user_indices)
bias_item = tf.nn.embedding_lookup(self.bi_var, self.item_indices)
embed_user = tf.nn.embedding_lookup(self.pu_var, self.user_indices)
embed_item = tf.nn.embedding_lookup(self.qi_var, self.item_indices)
self.output = bias_user + bias_item + tf.reduce_sum(
tf.multiply(embed_user, embed_item), axis=1)
def _build_dense(self):
self.dense_values = tf.placeholder(tf.float32,
shape=[None, self.dense_field_size])
dense_values_reshape = tf.reshape(self.dense_values,
[-1, self.dense_field_size, 1])
batch_size = tf.shape(self.dense_values)[0]
wide_dense_feat = tf.get_variable(name="wide_dense_feat",
shape=[self.dense_field_size],
initializer=tf_truncated_normal(
0.0, 0.01),
regularizer=self.reg)
deep_dense_feat = tf.get_variable(
name="deep_dense_feat",
shape=[self.dense_field_size, self.embed_size],
initializer=tf_truncated_normal(0.0, 0.01),
regularizer=self.reg)
wide_dense_embed = tf.tile(wide_dense_feat, [batch_size])
wide_dense_embed = tf.reshape(wide_dense_embed,
[-1, self.dense_field_size])
wide_dense_embed = tf.multiply(wide_dense_embed, self.dense_values)
deep_dense_embed = tf.expand_dims(deep_dense_feat, axis=0)
deep_dense_embed = tf.tile(deep_dense_embed, [batch_size, 1, 1])
deep_dense_embed = tf.multiply(deep_dense_embed, dense_values_reshape)
deep_dense_embed = tf.reshape(
deep_dense_embed, [-1, self.dense_field_size * self.embed_size])
self.wide_embed.append(wide_dense_embed)
self.deep_embed.append(deep_dense_embed)
def _build_sparse(self):
self.sparse_indices = tf.placeholder(
tf.int32, shape=[None, self.sparse_field_size])
linear_sparse_feat = tf.get_variable(name="linear_sparse_feat",
shape=[self.sparse_feature_size],
initializer=tf_truncated_normal(
0.0, 0.01),
regularizer=self.reg)
embed_sparse_feat = tf.get_variable(
name="embed_sparse_feat",
shape=[self.sparse_feature_size, self.embed_size],
initializer=tf_truncated_normal(0.0, 0.01),
regularizer=self.reg)
linear_sparse_embed = tf.nn.embedding_lookup( # B * F1
linear_sparse_feat, self.sparse_indices)
pairwise_sparse_embed = tf.nn.embedding_lookup( # B * F1 * K
embed_sparse_feat, self.sparse_indices)
deep_sparse_embed = tf.reshape(
pairwise_sparse_embed,
[-1, self.sparse_field_size * self.embed_size])
self.linear_embed.append(linear_sparse_embed)
self.pairwise_embed.append(pairwise_sparse_embed)
self.deep_embed.append(deep_sparse_embed)
def _build_dense(self):
self.dense_values = tf.placeholder(tf.float32,
shape=[None, self.dense_field_size])
dense_values_reshape = tf.reshape(self.dense_values,
[-1, self.dense_field_size, 1])
batch_size = tf.shape(self.dense_values)[0]
linear_dense_feat = tf.get_variable(name="linear_dense_feat",
shape=[self.dense_field_size],
initializer=tf_truncated_normal(
0.0, 0.03),
regularizer=self.reg)
pairwise_dense_feat = tf.get_variable(
name="pairwise_dense_feat",
shape=[self.dense_field_size, self.embed_size],
initializer=tf_truncated_normal(0.0, 0.03),
regularizer=self.reg)
# B * F2
linear_dense_embed = tf.tile(linear_dense_feat, [batch_size])
linear_dense_embed = tf.reshape(linear_dense_embed,
[-1, self.dense_field_size])
linear_dense_embed = tf.multiply(linear_dense_embed, self.dense_values)
pairwise_dense_embed = tf.expand_dims(pairwise_dense_feat, axis=0)
# B * F2 * K
pairwise_dense_embed = tf.tile(pairwise_dense_embed,
[batch_size, 1, 1])
pairwise_dense_embed = tf.multiply(pairwise_dense_embed,
dense_values_reshape)
self.linear_embed.append(linear_dense_embed)
self.pairwise_embed.append(pairwise_dense_embed)
def _build_model(self):
tf.set_random_seed(self.seed)
self.user_indices = tf.placeholder(tf.int32, shape=[None])
self.item_indices = tf.placeholder(tf.int32, shape=[None])
self.user_interacted_seq = tf.placeholder(
tf.int32, shape=[None, self.interaction_num])
self.user_interacted_len = tf.placeholder(tf.float32, shape=[None])
self.labels = tf.placeholder(tf.float32, shape=[None])
self.is_training = tf.placeholder_with_default(False, shape=[])
self.concat_embed = []
user_features = tf.get_variable(name="user_features",
shape=[self.n_users, self.embed_size],
initializer=tf_truncated_normal(
0.0, 0.01),
regularizer=self.reg)
item_features = tf.get_variable(
name="item_features",
shape=[self.n_items + 1, self.embed_size],
initializer=tf_truncated_normal(0.0, 0.01),
regularizer=self.reg)
user_embed = tf.nn.embedding_lookup(user_features, self.user_indices)
item_embed = tf.nn.embedding_lookup(item_features, self.item_indices)
# unknown items are padded to 0-vector
zero_padding_op = tf.scatter_update(
item_features, self.n_items,
tf.zeros([self.embed_size], dtype=tf.float32))
with tf.control_dependencies([zero_padding_op]):
multi_item_embed = tf.nn.embedding_lookup(
item_features, self.user_interacted_seq) # B * seq * K
pooled_embed = tf.div_no_nan(
tf.reduce_sum(multi_item_embed, axis=1),
tf.expand_dims(tf.sqrt(self.user_interacted_len), axis=1))
self.concat_embed.extend([user_embed, item_embed, pooled_embed])
if self.sparse:
self._build_sparse()
if self.dense:
self._build_dense()
concat_embed = tf.concat(self.concat_embed, axis=1)
mlp_layer = dense_nn(concat_embed,
self.hidden_units,
use_bn=self.use_bn,
dropout_rate=self.dropout_rate,
is_training=self.is_training)
self.output = tf.reshape(tf.layers.dense(inputs=mlp_layer, units=1),
[-1])
def _build_user_item(self):
self.user_indices = tf.placeholder(tf.int32, shape=[None])
self.item_indices = tf.placeholder(tf.int32, shape=[None])
linear_user_feat = tf.get_variable(name="linear_user_feat",
shape=[self.n_users, 1],
initializer=tf_truncated_normal(
0.0, 0.01),
regularizer=self.reg)
linear_item_feat = tf.get_variable(name="linear_item_feat",
shape=[self.n_items, 1],
initializer=tf_truncated_normal(
0.0, 0.01),
regularizer=self.reg)
embed_user_feat = tf.get_variable(
name="embed_user_feat",
shape=[self.n_users, self.embed_size],
initializer=tf_truncated_normal(0.0, 0.01),
regularizer=self.reg)
embed_item_feat = tf.get_variable(
name="embed_item_feat",
shape=[self.n_items, self.embed_size],
initializer=tf_truncated_normal(0.0, 0.01),
regularizer=self.reg)
linear_user_embed = tf.nn.embedding_lookup(linear_user_feat,
self.user_indices)
linear_item_embed = tf.nn.embedding_lookup(linear_item_feat,
self.item_indices)
self.linear_embed.extend([linear_user_embed, linear_item_embed])
pairwise_user_embed = tf.expand_dims(tf.nn.embedding_lookup(
embed_user_feat, self.user_indices),
axis=1)
pairwise_item_embed = tf.expand_dims(tf.nn.embedding_lookup(
embed_item_feat, self.item_indices),
axis=1)
self.pairwise_embed.extend([pairwise_user_embed, pairwise_item_embed])
deep_user_embed = tf.nn.embedding_lookup(embed_user_feat,
self.user_indices)
deep_item_embed = tf.nn.embedding_lookup(embed_item_feat,
self.item_indices)
self.deep_embed.extend([deep_user_embed, deep_item_embed])
def _build_model(self, sparse_implicit_interaction):
self.user_indices = tf.placeholder(tf.int32, shape=[None])
self.item_indices = tf.placeholder(tf.int32, shape=[None])
self.labels = tf.placeholder(tf.float32, shape=[None])
self.bu_var = tf.get_variable(name="bu_var",
shape=[self.n_users],
initializer=tf_zeros,
regularizer=self.reg)
self.bi_var = tf.get_variable(name="bi_var",
shape=[self.n_items],
initializer=tf_zeros,
regularizer=self.reg)
self.pu_var = tf.get_variable(name="pu_var",
shape=[self.n_users, self.embed_size],
initializer=tf_truncated_normal(
0.0, 0.03),
regularizer=self.reg)
self.qi_var = tf.get_variable(name="pi_var",
shape=[self.n_items, self.embed_size],
initializer=tf_truncated_normal(
0.0, 0.03),
regularizer=self.reg)
yj_var = tf.get_variable(name="yj_var",
shape=[self.n_items, self.embed_size],
initializer=tf_truncated_normal(0.0, 0.03),
regularizer=self.reg)
uj = tf.nn.safe_embedding_lookup_sparse(
yj_var,
sparse_implicit_interaction,
sparse_weights=None,
combiner="sqrtn",
default_id=None) # unknown user will return 0-vector
self.puj_var = self.pu_var + uj
bias_user = tf.nn.embedding_lookup(self.bu_var, self.user_indices)
bias_item = tf.nn.embedding_lookup(self.bi_var, self.item_indices)
embed_user = tf.nn.embedding_lookup(self.puj_var, self.user_indices)
embed_item = tf.nn.embedding_lookup(self.qi_var, self.item_indices)
self.output = bias_user + bias_item + tf.reduce_sum(
tf.multiply(embed_user, embed_item), axis=1)
def _build_train_ops(self, global_steps=None):
self.nce_weights = tf.get_variable(
name="nce_weights",
# n_classes, embed_size
shape=[self.n_items, self.user_vector_size],
initializer=tf_truncated_normal(0.0, 0.01),
regularizer=self.reg)
# we didn't add bias, since ANN can't be used with bias
self.nce_biases = tf.get_variable(name="nce_biases",
shape=[self.n_items],
initializer=tf_zeros,
regularizer=self.reg,
trainable=False)
if self.loss_type == "nce":
self.loss = tf.reduce_mean(
tf.nn.nce_loss(weights=self.nce_weights,
biases=self.nce_biases,
labels=tf.reshape(self.item_indices, [-1, 1]),
inputs=self.user_vector_repr,
num_sampled=self.num_neg,
num_classes=self.n_items,
num_true=1,
remove_accidental_hits=True,
partition_strategy="div"))
elif self.loss_type == "sampled_softmax":
self.loss = tf.reduce_mean(
tf.nn.sampled_softmax_loss(weights=self.nce_weights,
biases=self.nce_biases,
labels=tf.reshape(
self.item_indices, [-1, 1]),
inputs=self.user_vector_repr,
num_sampled=self.num_neg,
num_classes=self.n_items,
num_true=1,
remove_accidental_hits=True,
seed=self.seed,
partition_strategy="div"))
else:
raise ValueError("Loss type must either be 'nce' "
"or 'sampled_softmax")
if self.reg is not None:
reg_keys = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
total_loss = self.loss + tf.add_n(reg_keys)
else:
total_loss = self.loss
optimizer = tf.train.AdamOptimizer(self.lr)
optimizer_op = optimizer.minimize(total_loss, global_step=global_steps)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
self.training_op = tf.group([optimizer_op, update_ops])
self.sess.run(tf.global_variables_initializer())
def _build_model(self):
self.user_indices = tf.placeholder(tf.int32, shape=[None])
self.item_indices = tf.placeholder(tf.int32, shape=[None])
self.labels = tf.placeholder(tf.float32, shape=[None])
self.is_training = tf.placeholder_with_default(False, shape=[])
user_gmf = tf.get_variable(name="user_gmf",
shape=[self.n_users, self.embed_size],
initializer=tf_truncated_normal(0.0, 0.01),
regularizer=self.reg)
item_gmf = tf.get_variable(name="item_gmf",
shape=[self.n_items, self.embed_size],
initializer=tf_truncated_normal(0.0, 0.01),
regularizer=self.reg)
user_mlp = tf.get_variable(name="user_mlp",
shape=[self.n_users, self.embed_size],
initializer=tf_truncated_normal(0.0, 0.01),
regularizer=self.reg)
item_mlp = tf.get_variable(name="item_mlp",
shape=[self.n_items, self.embed_size],
initializer=tf_truncated_normal(0.0, 0.01),
regularizer=self.reg)
user_gmf_embed = tf.nn.embedding_lookup(user_gmf, self.user_indices)
item_gmf_embed = tf.nn.embedding_lookup(item_gmf, self.item_indices)
user_mlp_embed = tf.nn.embedding_lookup(user_mlp, self.user_indices)
item_mlp_embed = tf.nn.embedding_lookup(item_mlp, self.item_indices)
gmf_layer = tf.multiply(user_gmf_embed, item_gmf_embed)
mlp_input = tf.concat([user_mlp_embed, item_mlp_embed], axis=1)
mlp_layer = dense_nn(mlp_input,
self.hidden_units,
use_bn=self.use_bn,
dropout_rate=self.dropout_rate,
is_training=self.is_training)
concat_layer = tf.concat([gmf_layer, mlp_layer], axis=1)
self.output = tf.reshape(
tf.layers.dense(inputs=concat_layer, units=1), [-1])
def _build_variables(self):
self.user_feat = tf.get_variable(name="user_feat",
shape=[self.n_users, self.embed_size],
initializer=tf_truncated_normal(
0.0, 0.01),
regularizer=self.reg)
self.item_feat = tf.get_variable(name="item_feat",
shape=[self.n_items, self.embed_size],
initializer=tf_truncated_normal(
0.0, 0.01),
regularizer=self.reg)
if self.sparse:
self.sparse_feat = tf.get_variable(
name="sparse_feat",
shape=[self.sparse_feature_size, self.embed_size],
initializer=tf_truncated_normal(0.0, 0.01),
regularizer=self.reg)
if self.dense:
self.dense_feat = tf.get_variable(
name="dense_feat",
shape=[self.dense_field_size, self.embed_size],
initializer=tf_truncated_normal(0.0, 0.01),
regularizer=self.reg)
def _build_sparse(self):
self.sparse_indices = tf.placeholder(
tf.int32, shape=[None, self.sparse_field_size])
sparse_features = tf.get_variable(
name="sparse_features",
shape=[self.sparse_feature_size, self.embed_size],
initializer=tf_truncated_normal(0.0, 0.01),
regularizer=self.reg)
sparse_embed = tf.nn.embedding_lookup(sparse_features,
self.sparse_indices)
sparse_embed = tf.reshape(
sparse_embed, [-1, self.sparse_field_size * self.embed_size])
self.concat_embed.append(sparse_embed)
def _build_sparse(self):
self.sparse_indices = tf.placeholder(
tf.int32, shape=[None, self.sparse_field_size])
wide_sparse_feat = tf.get_variable(name="wide_sparse_feat",
shape=[self.sparse_feature_size],
initializer=tf_truncated_normal(
0.0, 0.01),
regularizer=self.reg)
deep_sparse_feat = tf.get_variable(
name="deep_sparse_feat",
shape=[self.sparse_feature_size, self.embed_size],
initializer=tf_truncated_normal(0.0, 0.01),
regularizer=self.reg)
wide_sparse_embed = tf.nn.embedding_lookup(wide_sparse_feat,
self.sparse_indices)
deep_sparse_embed = tf.nn.embedding_lookup(deep_sparse_feat,
self.sparse_indices)
deep_sparse_embed = tf.reshape(
deep_sparse_embed, [-1, self.sparse_field_size * self.embed_size])
self.wide_embed.append(wide_sparse_embed)
self.deep_embed.append(deep_sparse_embed)
def _build_user_item(self):
self.user_indices = tf.placeholder(tf.int32, shape=[None])
self.item_indices = tf.placeholder(tf.int32, shape=[None])
wide_user_feat = tf.get_variable(name="wide_user_feat",
shape=[self.n_users, 1],
initializer=tf_truncated_normal(
0.0, 0.01),
regularizer=self.reg)
wide_item_feat = tf.get_variable(name="wide_item_feat",
shape=[self.n_items, 1],
initializer=tf_truncated_normal(
0.0, 0.01),
regularizer=self.reg)
deep_user_feat = tf.get_variable(name="deep_user_feat",
shape=[self.n_users, self.embed_size],
initializer=tf_truncated_normal(
0.0, 0.01),
regularizer=self.reg)
deep_item_feat = tf.get_variable(name="deep_item_feat",
shape=[self.n_items, self.embed_size],
initializer=tf_truncated_normal(
0.0, 0.01),
regularizer=self.reg)
wide_user_embed = tf.nn.embedding_lookup(wide_user_feat,
self.user_indices)
wide_item_embed = tf.nn.embedding_lookup(wide_item_feat,
self.item_indices)
self.wide_embed.extend([wide_user_embed, wide_item_embed])
deep_user_embed = tf.nn.embedding_lookup(deep_user_feat,
self.user_indices)
deep_item_embed = tf.nn.embedding_lookup(deep_item_feat,
self.item_indices)
self.deep_embed.extend([deep_user_embed, deep_item_embed])
def _build_dense(self):
self.dense_values = tf.placeholder(tf.float32,
shape=[None, self.dense_field_size])
dense_values_reshape = tf.reshape(self.dense_values,
[-1, self.dense_field_size, 1])
batch_size = tf.shape(self.dense_values)[0]
dense_features = tf.get_variable(
name="dense_features",
shape=[self.dense_field_size, self.embed_size],
initializer=tf_truncated_normal(0.0, 0.01),
regularizer=self.reg)
dense_embed = tf.tile(dense_features, [batch_size, 1])
dense_embed = tf.reshape(dense_embed,
[-1, self.dense_field_size, self.embed_size])
dense_embed = tf.multiply(dense_embed, dense_values_reshape)
dense_embed = tf.reshape(dense_embed,
[-1, self.dense_field_size * self.embed_size])
self.concat_embed.append(dense_embed)
def _build_item_interaction(self):
self.item_interaction_indices = tf.placeholder(tf.int64,
shape=[None, 2])
self.item_interaction_values = tf.placeholder(tf.int32, shape=[None])
self.modified_batch_size = tf.placeholder(tf.int32, shape=[])
item_interaction_features = tf.get_variable(
name="item_interaction_features",
shape=[self.n_items, self.embed_size],
initializer=tf_truncated_normal(0.0, 0.01),
regularizer=self.reg)
sparse_item_interaction = tf.SparseTensor(
self.item_interaction_indices, self.item_interaction_values,
[self.modified_batch_size, self.n_items])
pooled_embed = tf.nn.safe_embedding_lookup_sparse(
item_interaction_features,
sparse_item_interaction,
sparse_weights=None,
combiner="sqrtn",
default_id=None) # unknown user will return 0-vector
self.concat_embed.append(pooled_embed)