"""Implementation of NCF."""
from ..bases import ModelMeta, TfBase
from ..layers import dense_nn, embedding_lookup, tf_dense
from ..tfops import dropout_config, reg_config, tf
from ..torchops import hidden_units_config
[docs]class NCF(TfBase, metaclass=ModelMeta):
"""*Neural Collaborative Filtering* algorithm.
Parameters
----------
task : {'rating', 'ranking'}
Recommendation task. See :ref:`Task`.
data_info : :class:`~libreco.data.DataInfo` object
Object that contains useful information for training and inference.
loss_type : {'cross_entropy', 'focal'}, default: 'cross_entropy'
Loss for model training.
embed_size: int, default: 16
Vector size of embeddings.
n_epochs: int, default: 10
Number of epochs for training.
lr : float, default 0.001
Learning rate for training.
lr_decay : bool, default: False
Whether to use learning rate decay.
epsilon : float, default: 1e-5
A small constant added to the denominator to improve numerical stability in
Adam optimizer.
According to the `official comment <https://github.com/tensorflow/tensorflow/blob/v1.15.0/tensorflow/python/training/adam.py#L64>`_,
default value of `1e-8` for `epsilon` is generally not good, so here we choose `1e-5`.
Users can try tuning this hyperparameter if the training is unstable.
reg : float or None, default: None
Regularization parameter, must be non-negative or None.
batch_size : int, default: 256
Batch size for training.
sampler : {'random', 'unconsumed', 'popular'}, default: 'random'
Negative sampling strategy.
- ``'random'`` means random sampling.
- ``'unconsumed'`` samples items that the target user did not consume before.
- ``'popular'`` has a higher probability to sample popular items as negative samples.
.. versionadded:: 1.1.0
num_neg : int, default: 1
Number of negative samples for each positive sample, only used in `ranking` task.
use_bn : bool, default: True
Whether to use batch normalization.
dropout_rate : float or None, default: None
Probability of an element to be zeroed. If it is None, dropout is not used.
hidden_units : int, list of int or tuple of (int,), default: (128, 64, 32)
Number of layers and corresponding layer size in MLP.
.. versionchanged:: 1.0.0
Accept type of ``int``, ``list`` or ``tuple``, instead of ``str``.
seed : int, default: 42
Random seed.
lower_upper_bound : tuple or None, default: None
Lower and upper score bound for `rating` task.
tf_sess_config : dict or None, default: None
Optional TensorFlow session config, see `ConfigProto options
<https://github.com/tensorflow/tensorflow/blob/v2.10.0/tensorflow/core/protobuf/config.proto#L431>`_.
References
----------
*Xiangnan He et al.* `Neural Collaborative Filtering
<https://arxiv.org/pdf/1708.05031.pdf>`_.
"""
user_variables = ("embedding/user_embeds_var",)
item_variables = ("embedding/item_embeds_var",)
def __init__(
self,
task,
data_info,
loss_type="cross_entropy",
embed_size=16,
n_epochs=20,
lr=0.01,
lr_decay=False,
epsilon=1e-5,
reg=None,
batch_size=256,
sampler="random",
num_neg=1,
use_bn=True,
dropout_rate=None,
hidden_units=(128, 64, 32),
seed=42,
lower_upper_bound=None,
tf_sess_config=None,
):
super().__init__(task, data_info, lower_upper_bound, tf_sess_config)
self.all_args = locals()
self.loss_type = loss_type
self.embed_size = embed_size
self.n_epochs = n_epochs
self.lr = lr
self.lr_decay = lr_decay
self.epsilon = epsilon
self.reg = reg_config(reg)
self.batch_size = batch_size
self.sampler = sampler
self.num_neg = num_neg
self.use_bn = use_bn
self.dropout_rate = dropout_config(dropout_rate)
self.hidden_units = hidden_units_config(hidden_units)
self.seed = seed
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.labels = tf.placeholder(tf.float32, shape=[None])
self.is_training = tf.placeholder_with_default(False, shape=[])
user_embeds = embedding_lookup(
indices=self.user_indices,
var_name="user_embeds_var",
var_shape=(self.n_users + 1, self.embed_size),
initializer=tf.glorot_uniform_initializer(),
regularizer=self.reg,
)
item_embeds = embedding_lookup(
indices=self.item_indices,
var_name="item_embeds_var",
var_shape=(self.n_items + 1, self.embed_size),
initializer=tf.glorot_uniform_initializer(),
regularizer=self.reg,
)
gmf_layer = tf.multiply(user_embeds, item_embeds)
mlp_input = tf.concat([user_embeds, item_embeds], 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_dense(units=1)(concat_layer), [-1])
self.serving_topk = self.build_topk(self.output)