Source code for libreco.algorithms.youtube_retrieval

"""Implementation of YouTubeRetrieval."""
import numpy as np

from ..bases import EmbedBase, ModelMeta
from ..tfops import (
    dense_nn,
    dropout_config,
    multi_sparse_combine_embedding,
    reg_config,
    sess_config,
    tf,
)
from ..torchops import hidden_units_config
from ..utils.misc import count_params
from ..utils.validate import (
    check_dense_values,
    check_seq_mode,
    check_multi_sparse,
    check_sparse_indices,
    dense_field_size,
    sparse_feat_size,
    sparse_field_size,
    true_sparse_field_size,
)


[docs]class YouTubeRetrieval(EmbedBase, metaclass=ModelMeta, backend="tensorflow"): """*YouTubeRetrieval* algorithm. .. NOTE:: The algorithm implemented mainly corresponds to the candidate generation phase based on the original paper. .. WARNING:: YouTubeRetrieval can only be used in `ranking` task. Parameters ---------- task : {'ranking'} Recommendation task. See :ref:`Task`. data_info : :class:`~libreco.data.DataInfo` object Object that contains useful information for training and inference. loss_type : {'sampled_softmax', 'nce'}, default: 'sampled_softmax' 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. 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) Number of layers and corresponding layer size in MLP. .. versionchanged:: 1.0.0 Accept type of ``int``, ``list`` or ``tuple``, instead of ``str``. num_sampled_per_batch : int or None, default: None Number of negative samples in a batch. If None, it is set to `batch_size`. sampler : str, default: 'uniform' Negative Sampling strategy. 'uniform' will use uniform sampler, and setting to other value will use `log_uniform_candidate_sampler` in TensorFlow. In recommendation scenarios the uniform sampler is generally preferred. recent_num : int or None, default: 10 Number of recent items to use in user behavior sequence. random_num : int or None, default: None Number of random sampled items to use in user behavior sequence. If `recent_num` is not None, `random_num` is not considered. multi_sparse_combiner : {'normal', 'mean', 'sum', 'sqrtn'}, default: 'sqrtn' Options for combining `multi_sparse` features. 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 ---------- *Paul Covington et al.* `Deep Neural Networks for YouTube Recommendations <https://static.googleusercontent.com/media/research.google.com/zh-CN//pubs/archive/45530.pdf>`_. """ item_variables = ["item_interaction_features", "nce_weights", "nce_biases"] sparse_variables = ["sparse_features"] dense_variables = ["dense_features"] def __init__( self, task="ranking", data_info=None, loss_type="sampled_softmax", embed_size=16, n_epochs=20, lr=0.001, lr_decay=False, epsilon=1e-5, reg=None, batch_size=256, use_bn=True, dropout_rate=None, hidden_units=(128, 64), num_sampled_per_batch=None, sampler="uniform", recent_num=10, random_num=None, multi_sparse_combiner="sqrtn", seed=42, lower_upper_bound=None, tf_sess_config=None, ): super().__init__(task, data_info, embed_size, lower_upper_bound) assert task == "ranking", "YouTube-type models is only suitable for ranking" if len(data_info.item_col) > 0: raise ValueError("The `YouTuBeRetrieval` model assumes no item features.") self.all_args = locals() self.loss_type = loss_type self.n_epochs = n_epochs self.lr = lr self.lr_decay = lr_decay self.epsilon = epsilon self.hidden_units = hidden_units_config(hidden_units) + [self.embed_size] self.reg = reg_config(reg) self.batch_size = batch_size self.use_bn = use_bn self.dropout_rate = dropout_config(dropout_rate) self.num_sampled_per_batch = num_sampled_per_batch self.sampler = sampler self.seed = seed self.sess = sess_config(tf_sess_config) self.seq_mode, self.max_seq_len = check_seq_mode(recent_num, random_num) self.recent_seq_indices, self.recent_seq_values = self._set_recent_seqs() self.sparse = check_sparse_indices(data_info) self.dense = check_dense_values(data_info) if self.sparse: self.sparse_feature_size = sparse_feat_size(data_info) self.sparse_field_size = sparse_field_size(data_info) self.multi_sparse_combiner = check_multi_sparse( data_info, multi_sparse_combiner ) self.true_sparse_field_size = true_sparse_field_size( data_info, self.sparse_field_size, self.multi_sparse_combiner ) if self.dense: self.dense_field_size = dense_field_size(data_info) def build_model(self): tf.set_random_seed(self.seed) # item_indices actually serve as labels in YouTuBeRetrieval model self.item_indices = tf.placeholder(tf.int64, shape=[None]) self.is_training = tf.placeholder_with_default(False, shape=[]) self.concat_embed = [] self._build_item_interaction() self._build_variables() if self.sparse: self._build_sparse() if self.dense: self._build_dense() concat_features = tf.concat(self.concat_embed, axis=1) self.user_vector_repr = dense_nn( concat_features, self.hidden_units, use_bn=self.use_bn, dropout_rate=self.dropout_rate, is_training=self.is_training, ) count_params() 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.glorot_uniform_initializer(), 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) 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.glorot_uniform_initializer(), regularizer=self.reg, ) if self.data_info.multi_sparse_combine_info and self.multi_sparse_combiner in ( "sum", "mean", "sqrtn", ): sparse_embed = multi_sparse_combine_embedding( self.data_info, sparse_features, self.sparse_indices, self.multi_sparse_combiner, self.embed_size, ) else: sparse_embed = tf.nn.embedding_lookup(sparse_features, self.sparse_indices) sparse_embed = tf.reshape( sparse_embed, [-1, self.true_sparse_field_size * self.embed_size] ) self.concat_embed.append(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] dense_features = tf.get_variable( name="dense_features", shape=[self.dense_field_size, self.embed_size], initializer=tf.glorot_uniform_initializer(), regularizer=self.reg, ) dense_embed = tf.expand_dims(dense_features, axis=0) # B * F2 * K dense_embed = tf.tile(dense_embed, [batch_size, 1, 1]) 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_variables(self): self.nce_weights = tf.get_variable( name="nce_weights", # n_classes, embed_size shape=[self.n_items, self.embed_size], initializer=tf.glorot_uniform_initializer(), regularizer=self.reg, ) self.nce_biases = tf.get_variable( name="nce_biases", shape=[self.n_items], initializer=tf.zeros_initializer(), regularizer=self.reg, trainable=True, ) def _set_recent_seqs(self): interacted_indices = [] interacted_items = [] for u in range(self.n_users): u_consumed_items = self.user_consumed[u] u_items_len = len(u_consumed_items) if u_items_len < self.max_seq_len: interacted_indices.extend([u] * u_items_len) interacted_items.extend(u_consumed_items) else: interacted_indices.extend([u] * self.max_seq_len) interacted_items.extend(u_consumed_items[-self.max_seq_len:]) interacted_indices = np.array(interacted_indices).reshape(-1, 1) indices = np.concatenate( [interacted_indices, np.zeros_like(interacted_indices)], axis=1 ) return indices, interacted_items def set_embeddings(self): feed_dict = { self.item_interaction_indices: self.recent_seq_indices, self.item_interaction_values: self.recent_seq_values, self.modified_batch_size: self.n_users, self.is_training: False, } if self.sparse: # remove oov user_sparse_indices = self.data_info.user_sparse_unique[:-1] feed_dict.update({self.sparse_indices: user_sparse_indices}) if self.dense: user_dense_values = self.data_info.user_dense_unique[:-1] feed_dict.update({self.dense_values: user_dense_values}) user_vector = self.sess.run(self.user_vector_repr, feed_dict) item_weights = self.sess.run(self.nce_weights) item_biases = self.sess.run(self.nce_biases) user_bias = np.ones([len(user_vector), 1], dtype=user_vector.dtype) item_bias = item_biases[:, None] self.user_embed = np.hstack([user_vector, user_bias]) self.item_embed = np.hstack([item_weights, item_bias])