def __init__(self, feature_metas, embedding_dim=8, hidden_units=(128, 64), use_dnn=True, linear_bias=False):
super(AFM, self).__init__()
self.feature_metas = feature_metas
self.linear_model = LinearModel(feature_metas, linear_bias=linear_bias)
self.embeddings = GroupEmbedding(feature_metas, embedding_dim)
self.att_pool = AttentionPooling(feature_metas, embedding_dim=embedding_dim, attention_factor=4)
self.dnn = MLP([embedding_dim] + list(hidden_units), dropout_prob=0.5, use_bn=True) if use_dnn else None
self.pred = Prediction()
def __init__(self, feature_metas, embedding_dim=8, hidden_units=(128, 64), linear_bias=False):
super(WideDeep, self).__init__()
self.feature_metas = feature_metas
self.linear_model = LinearModel(feature_metas, linear_bias=linear_bias)
self.embeddings = GroupEmbedding(feature_metas, embedding_dim)
self.dnn = MLP([len(feature_metas) * embedding_dim] + list(hidden_units),
activation='relu', dropout_prob=0.5, use_bn=True)
self.pred = Prediction()
def __init__(self,
feature_metas,
embedding_dim=8,
hidden_units=(256, 64),
linear_bias=False,
dense_bias=False):
super(xDeepFM, self).__init__()
self.feature_metas = feature_metas
self.linear_model = LinearModel(feature_metas, linear_bias=linear_bias)
self.embeddings = GroupEmbedding(feature_metas, embedding_dim)
self.cin = CIN(input_dims=(len(feature_metas), embedding_dim))
self.dnn = MLP(hidden_units=[len(feature_metas) * embedding_dim] +
list(hidden_units),
use_bn=True)
self.pred = Prediction()
def __init__(self,
feature_metas,
embedding_dim=8,
hidden_units=(256, 64),
linear_bias=False,
dense_bias=False):
super(NFFM, self).__init__()
self.feature_metas = feature_metas
self.linear_model = LinearModel(feature_metas, linear_bias=linear_bias)
self.embeddings = GroupEmbedding(feature_metas, embedding_dim)
# self.fm = FM()
self.dnn = MLP(hidden_units=[len(feature_metas) * embedding_dim] +
list(hidden_units),
bias=dense_bias)
self.pred = Prediction()
def __init__(self,
feature_metas,
embedding_dim=8,
hidden_units=(128, 64),
bi_dropout_prob=0.,
linear_bias=False):
super(NFM, self).__init__()
self.feature_metas = feature_metas
self.linear_model = LinearModel(feature_metas, linear_bias=linear_bias)
self.embeddings = GroupEmbedding(feature_metas, embedding_dim)
self.bi_interaction = BiInteractionPooling(bi_dropout_prob)
self.dnn = MLP([embedding_dim] + list(hidden_units),
activation='relu',
use_bn=True)
self.pred = Prediction(task='binary')
def __init__(self,
feature_metas,
embedding_dim=8,
hidden_units=(128, 64),
linear_bias=False):
super(DCN, self).__init__()
self.feature_metas = feature_metas
self.linear_model = LinearModel(feature_metas, linear_bias=linear_bias)
self.embeddings = GroupEmbedding(feature_metas, embedding_dim)
self.cross = CrossNet(len(feature_metas), embedding_dim, layer_num=3)
self.dnn = MLP([len(feature_metas) * embedding_dim] +
list(hidden_units),
activation='relu',
use_bn=True)
self.pred = Prediction()
def __init__(self,
feature_metas,
embedding_dim=8,
hidden_units=(128, 64),
product_mode='outer',
linear_bias=False):
super(PNN, self).__init__()
# mode : 'inner' or 'outer'
self.feature_metas = feature_metas
self.linear_model = LinearModel(feature_metas, linear_bias)
self.embeddings = GroupEmbedding(feature_metas, embedding_dim)
if product_mode == 'inner':
self.product = InnerProduct(feature_metas, embedding_dim)
else:
self.product = OuterProduct(feature_metas, embedding_dim)
m = (len(feature_metas) + 1) * len(
feature_metas) // 2 # TODO, require_logits affect this shape
self.dnn = MLP([m] + list(hidden_units), use_bn=True)
self.pred = Prediction()
def __init__(self, feature_metas, embedding_dim=8, linear_bias=False):
super(FM, self).__init__()
self.linear_model = LinearModel(feature_metas, linear_bias=linear_bias)
self.embeddings = GroupEmbedding(feature_metas, embedding_dim)
self.pred = Prediction()