class UCML(Model):
def __init__(self,
dim_user_embed,
dim_item_embed,
total_users,
total_items,
margin=0.5):
super(UCML, self).__init__()
self.user_latent_factor = LatentFactor(num_instances=total_users,
dim=dim_user_embed,
name='user_latent_factor')
self.item_latent_factor = LatentFactor(num_instances=total_items,
dim=dim_item_embed,
name='item_latent_factor')
self.item_bias = LatentFactor(num_instances=total_items,
dim=1,
name='item_bias')
self.margin = margin
def call(self, user_id, p_item_id, n_item_id):
user_vec = self.user_latent_factor(user_id)
p_item_vec = self.item_latent_factor(p_item_id)
p_item_bias = self.item_bias(p_item_id)
n_item_vec = self.item_latent_factor(n_item_id)
n_item_bias = self.item_bias(n_item_id)
l2_user_pos = tf.math.reduce_sum(tf.math.square(user_vec - p_item_vec),
axis=-1,
keepdims=True)
l2_user_neg = tf.math.reduce_sum(tf.math.square(user_vec - n_item_vec),
axis=-1,
keepdims=True)
pos_score = (-l2_user_pos) + p_item_bias
neg_score = (-l2_user_neg) + n_item_bias
diff = pos_score - neg_score
loss = tf.reduce_sum(tf.maximum(self.margin - diff, 0))
l2_loss = tf.nn.l2_loss(user_vec) + tf.nn.l2_loss(
p_item_vec) + tf.nn.l2_loss(n_item_vec)
return loss, l2_loss
def censor_vec(self, user_id, p_item_id, n_item_id):
return self.user_latent_factor.censor(user_id), \
self.item_latent_factor.censor(p_item_id), \
self.item_latent_factor.censor(n_item_id)
def inference(self, user_id):
user_vec = self.user_latent_factor(user_id)
return -tf.math.reduce_sum(tf.math.square(
tf.expand_dims(user_vec, axis=1) -
self.item_latent_factor.variables[0]),
axis=-1,
keepdims=False) + tf.reshape(
self.item_bias.variables[0], [-1])
def __init__(self, dim_user_embed, dim_item_embed, total_users, total_items):
super(BPR, self).__init__()
self.user_latent_factor = LatentFactor(num_instances=total_users,
dim=dim_user_embed,
name='user_latent_factor')
self.item_latent_factor = LatentFactor(num_instances=total_items,
dim=dim_item_embed,
name='item_latent_factor')
self.item_bias = LatentFactor(num_instances=total_items,
dim=1,
name='item_bias')
self.pairwise_log_loss = PairwiseLogLoss()
def __init__(self, dim_user_embed, dim_item_embed, total_users, total_items, a=1.0, b=1.0):
super(WRMF, self).__init__()
self.user_latent_factor = LatentFactor(num_instances=total_users,
dim=dim_user_embed,
name='user_latent_factor')
self.item_latent_factor = LatentFactor(num_instances=total_items,
dim=dim_item_embed,
name='item_latent_factor')
self.item_bias = LatentFactor(num_instances=total_items,
dim=1,
name='item_bias')
self.pointwise_mse_loss = PointwiseMSELoss(a=a, b=b)
def __init__(self, dim_user_embed, dim_item_embed, total_users, total_items):
super(GMF, self).__init__()
self.user_latent_factor = LatentFactor(num_instances=total_users,
dim=dim_user_embed,
name='user_latent_factor')
self.item_latent_factor = LatentFactor(num_instances=total_items,
dim=dim_item_embed,
name='item_latent_factor')
self.item_bias = LatentFactor(num_instances=total_items,
dim=1,
name='item_bias')
self.mlp = MLP(units_list=[1], use_bias=False)
self._bce = tf.keras.losses.BinaryCrossentropy(from_logits=True)
def __init__(self,
dim_user_embed,
dim_item_embed,
total_users,
total_items,
margin=0.5):
super(UCML, self).__init__()
self.user_latent_factor = LatentFactor(num_instances=total_users,
dim=dim_user_embed,
name='user_latent_factor')
self.item_latent_factor = LatentFactor(num_instances=total_items,
dim=dim_item_embed,
name='item_latent_factor')
self.item_bias = LatentFactor(num_instances=total_items,
dim=1,
name='item_bias')
self.margin = margin
def __init__(self,
m_spa,
ln_emb,
ln_bot,
ln_top,
arch_interaction_op='dot',
arch_interaction_itself=False,
sigmoid_bot=False,
sigmoid_top=True,
loss_func='mse',
loss_threshold=0.0):
'''
m_spa: the dimensionality of sparse feature embeddings
ln_emb: the size of sparse feature embeddings (num_instances)
ln_bot: the size of the bottom MLP
ln_top: the size of the top MLP
'''
super(DLRM, self).__init__()
self._loss_threshold = loss_threshold
self._loss_func = loss_func
self._latent_factors = [
LatentFactor(num_instances=num, dim=m_spa) for num in ln_emb
]
self._mlp_bot = MLP(
units_list=ln_bot,
out_activation='sigmoid' if sigmoid_bot else 'relu')
self._mlp_top = MLP(
units_list=ln_top,
out_activation='sigmoid' if sigmoid_top else 'relu')
self._dot_interaction = None
if arch_interaction_op == 'dot':
self._dot_interaction = SecondOrderFeatureInteraction(
self_interaction=arch_interaction_itself)
elif self._arch_interaction_op != 'cat':
sys.exit("ERROR: arch_interaction_op=" +
self._arch_interaction_op + " is not supported")
if loss_func == 'mse':
self._loss = tf.keras.losses.MeanSquaredError()
elif loss_func == 'bce':
self._loss = tf.keras.losses.BinaryCrossentropy()
else:
sys.exit("ERROR: loss_func=" + loss_func + " is not supported")