svd.py

"""
Singular Value Decomposition model
"""

import tensorflow as tf


def svd(users_number: int, items_number: int):

    """
    :param users_number: int
    :param items_number: int
    :return: SVD model
    """

    latent_dim, max_rating, min_rating = 10, 5, 1

    # define placeholder.
    user_id_input = tf.keras.layers.Input(shape=[1], name='user')
    item_id_input = tf.keras.layers.Input(shape=[1], name='item')

    # define embedding size and layers.

    user_embedding = tf.keras.layers.Embedding(
        output_dim=latent_dim,
        input_dim=users_number + 1,
        input_length=1,
        name='user_embedding',
        embeddings_regularizer=tf.keras.regularizers.l2())(user_id_input)

    item_embedding = tf.keras.layers.Embedding(
        output_dim=latent_dim,
        input_dim=items_number + 1,
        input_length=1,
        name='item_embedding',
        embeddings_regularizer=tf.keras.regularizers.l2())(item_id_input)

    user_bias = tf.keras.layers.Embedding(input_dim=users_number + 1, output_dim=1)(user_id_input)
    item_bias = tf.keras.layers.Embedding(input_dim=items_number + 1, output_dim=1)(item_id_input)

    user_vecs = tf.keras.layers.Reshape([latent_dim])(user_embedding)
    item_vecs = tf.keras.layers.Reshape([latent_dim])(item_embedding)

    # The prediction, which we calculate the loss function with ground truth and optimize.
    y_hat = tf.keras.layers.Dot(1, normalize=False)([user_vecs, item_vecs])

    # add tf.keras.backend.constant(np.mean(targets), shape=[])
    y_hat = tf.keras.layers.Add()([y_hat, user_bias, item_bias])

    output = tf.keras.layers.Activation('relu')(y_hat)
    output = tf.keras.layers.Lambda(lambda x: x * (max_rating - min_rating) + min_rating)(output)

    model = tf.keras.models.Model(inputs=[user_id_input, item_id_input], outputs=output)

    return model