def __init__(self, feature_cards, factor_dim, hidden_sizes, dropout_rate=.1, name='fnfm'):
super(FieldAwareNeuralFactorizationMachine, self).__init__(name=name)
self.num_features = len(feature_cards)
self.embeddings = FieldAwareEmbedFeatures(feature_cards, factor_dim,
name=name + '/field_aware_feature_embedding')
self.linear = LinearModel(feature_cards, name=name + '/linear_model')
self.nn = FullyConnectedNetwork(units=hidden_sizes, dropout_rate=dropout_rate, name=name + '/fcn')
def __init__(self, feature_cards, factor_dim, name='ffm'):
super(FieldAwareFactorizationMachine, self).__init__(name=name)
self.factor_dim = factor_dim
self.embeddings = FieldAwareEmbedFeatures(
feature_cards,
factor_dim,
name=name + '/field_aware_feature_embeddings')
self.linear = LinearModel(feature_cards, name=name + '/linear_model')
def __init__(self,
feature_cards,
factor_dim,
name='factorization_machine'):
super(FactorizationMachine, self).__init__(name=name)
self.embedding = EmbedFeatures(feature_cards,
factor_dim,
name=name + '/feature_embedding')
self.linear = LinearModel(feature_cards, name=name + '/linear_model')
def __init__(self, feature_cards, factor_dim, n_heads, n_attentions, hidden_sizes, dropout_rate=.1, name='deepfm'):
super(AutomaticFeatureInteraction, self).__init__(name=name)
self.linear = LinearModel(feature_cards, name=name + '/linear_model')
self.embedding = EmbedFeatures(feature_cards, factor_dim, name=name + '/feature_embedding')
self.flatten = tf.keras.layers.Flatten(data_format='channels_first')
self.nn = FullyConnectedNetwork(units=hidden_sizes, dropout_rate=dropout_rate, name=name + '/fcn')
self.attns = [
MultiHeadAttention(n_heads, factor_dim, dropout_rate=dropout_rate, name=name + '/mhattn{}'.format(i))
for i in range(n_attentions)
]
self.attn_out = tf.keras.layers.Dense(1)
def __init__(self, feature_cards, factor_dim, attention_size, name='afm'):
super(AttentionalFactorizationMachine, self).__init__(name=name)
self.num_features = len(feature_cards)
self.factor_dim = factor_dim
self.linear = LinearModel(feature_cards, name=name + '/linear_model')
self.embedding = EmbedFeatures(feature_cards, factor_dim, name=name + '/feature_embedding')
self.attention = tf.keras.Sequential([
tf.keras.layers.Dense(units=attention_size, name=name + '/attention_hidden'),
tf.keras.layers.ReLU(name=name + '/attention_activ'),
tf.keras.layers.Dense(units=1, name=name + '/attention_logits'),
tf.keras.layers.Softmax(axis=1, name=name + '/attention_score')
])
def __init__(self,
feature_cards,
factor_dim,
hidden_sizes,
dropout_rate=.1,
name='neural_factorization_machine'):
super(NeuralFactorizationMachine, self).__init__(name=name)
self.embedding = EmbedFeatures(feature_cards,
factor_dim,
name=name + '/feature_embedding')
self.linear = LinearModel(feature_cards, name=name + '/linear_model')
self.nn = FullyConnectedNetwork(units=hidden_sizes,
dropout_rate=dropout_rate,
name=name + '/fcn')
def __init__(self,
feature_cards,
factor_dim,
hidden_sizes,
dropout_rate=.1,
name='deepfm'):
super(DeepFM, self).__init__(name=name)
self.linear = LinearModel(feature_cards, name=name + '/linear_model')
self.embedding = EmbedFeatures(feature_cards,
factor_dim,
name=name + '/feature_embedding')
self.flatten = tf.keras.layers.Flatten(data_format='channels_first')
self.nn = FullyConnectedNetwork(units=hidden_sizes,
dropout_rate=dropout_rate,
name=name + '/fcn')
def __init__(self,
feature_cards,
factor_dim,
fnn_hidden_sizes,
cin_hidden_sizes,
dropout_rate=.1,
split=True,
name='xdeepfm'):
super(ExtremeDeepFactorizationMachine, self).__init__(name=name)
self.linear = LinearModel(feature_cards, name=name + '/linear_model')
self.embedding = EmbedFeatures(feature_cards,
factor_dim,
name=name + '/feature_embedding')
self.flatten = tf.keras.layers.Flatten(data_format='channels_first')
self.nn = FullyConnectedNetwork(units=fnn_hidden_sizes,
dropout_rate=dropout_rate,
name=name + '/fcn')
self.cin = CompressedInteractionNetwork(cin_hidden_sizes,
split,
name=name + '/cin')