@TOC
下图为NeutralCF的模型结构图，总共两个分支，第一个分支为GML，第二个为MLP，GML通路将两个特征的Embedding向量进行内积操作，MLP将两个特征的Embedding的向量进行拼接，然后使用多层感知机进行传播，然后将两个通路输出的向量进行拼接，导入全连接层（输出层），输出Score。
一、导包
import tensorflow as tf
from tensorflow.keras.layers import *
from tensorflow.keras.models import *
from tensorflow.keras.utils import plot_model
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import MinMaxScaler, LabelEncoder
import itertools
import pandas as pd
import numpy as np
from tqdm import tqdm
from collections import namedtuple
import warnings
warnings.filterwarnings("ignore")
二、读取数据
# 读取数据，NCF使用的特征只有user_id和item_id
rnames = ['user_id','movie_id','rating','timestamp']
data = pd.read_csv('./data/ml-1m/ratings.dat', sep='::', engine='python', names=rnames)
三、特征编码处理
lbe = LabelEncoder()
data['user_id'] = lbe.fit_transform(data['user_id'])
data['movie_id'] = lbe.fit_transform(data['movie_id'])
train_data = data[['user_id', 'movie_id']]
train_data['label'] = data['rating']
四、使用具名元组为特征进行处理
SparseFeat = namedtuple('SparseFeat', ['name', 'vocabulary_size', 'embedding_dim'])
DenseFeat = namedtuple('DenseFeat', ['name', 'dimension'])
dnn_features_columns = [SparseFeat('user_id', train_data['user_id'].nunique(), 8),
SparseFeat('movie_id', train_data['movie_id'].nunique(), 8)]
五、构建模型
5.1 输入层
def build_input_layers(dnn_features_columns):
dense_input_dict, sparse_input_dict = {}, {}
for f in dnn_features_columns:
if isinstance(f, SparseFeat):
sparse_input_dict[f.name] = Input(shape=(1), name=f.name)
elif isinstance(f, DenseFeat):
dense_input_dict[f.name] = Input(shape=(f.dimension), name=f.name)
return dense_input_dict, sparse_input_dict
5.2 Embedding层
def build_embedding_layers(dnn_features_columns, sparse_input_dict, prefix="", is_linear=True):
embedding_layers_dict = {}
sparse_feature_columns = list(filter(lambda x: isinstance(x, SparseFeat), dnn_features_columns)) if dnn_features_columns else []
if is_linear:
for f in sparse_feature_columns:
embedding_layers_dict[f.name] = Embedding(f.vocabulary_size + 1, 1, name= prefix + '_1d_emb_' + + f.name)
else:
for f in sparse_feature_columns:
embedding_layers_dict[f.name] = Embedding(f.vocabulary_size + 1, f.embedding_dim, name=prefix + '_kd_emb_' + f.name)
return embedding_layers_dict
5.3 GML
def build_gml_layers(gml_user_embedding, gml_movie_embedding):
return Multiply()([gml_user_embedding, gml_movie_embedding])
5.4 MLP
def build_mlp_layers(mlp_input, units=(32, 16)):
for out_dim in units:
mlp_input = Dense(out_dim)(mlp_input)
return mlp_input
5.5 输出层
def bulid_output_layers(concat_output):
return Dense(1)(concat_output)
5.6 构建模型
def NCF(dnn_features_columns):
# 1. 获取字典输入层，键为列名，值为对应的Input
_, sparse_input_dict = build_input_layers(dnn_features_columns)
# 2. 获取真实输入层，使用列表存储每个列的Input
input_layers = list(sparse_input_dict.values())
# 3. 将SparseFeature进行Embedding，有两路，分别是GML和MLP
embedding_gml_dict = build_embedding_layers(dnn_features_columns, sparse_input_dict, prefix="GML", is_linear=False)
embedding_mlp_dict = build_embedding_layers(dnn_features_columns, sparse_input_dict, prefix="MLP", is_linear=False)
# 4. 将Embedding后的特征进行展开，因为Embedding后为(?,1,8)
gml_user_embedding = Flatten()(embedding_gml_dict['user_id'](sparse_input_dict['user_id']))
gml_movie_embedding = Flatten()(embedding_gml_dict['movie_id'](sparse_input_dict['movie_id']))
mlp_user_embedding = Flatten()(embedding_mlp_dict['user_id'](sparse_input_dict['user_id']))
mlp_movie_embedding = Flatten()(embedding_mlp_dict['movie_id'](sparse_input_dict['movie_id']))
# 5. 进行GML，就是展开后的特征进行内积
gml_output = build_gml_layers(gml_user_embedding, gml_movie_embedding)
# gml_output = tf.multiply(gml_movie_embedding, gml_user_embedding)
# gml_output = Multiply()([gml_user_embedding, gml_movie_embedding])
# 6. 进行MLP，将特征进行连接，传入MLP层
mlp_input = Concatenate(axis=1)([mlp_user_embedding, mlp_movie_embedding])
mlp_output = build_mlp_layers(mlp_input, (32, 16))
# 7. 将GML和MLP层的输出进行连接
concat_output = Concatenate(axis=1)([gml_output, mlp_output])
# 8.传入到输出层中，获取评分
output_layers = bulid_output_layers(concat_output)
# 构建模型
model = Model(input_layers, output_layers)
return model
六、运转模型
history = NCF(dnn_features_columns)
# 编译模型
history.compile(optimizer="adam",
loss="mse",
metrics=['mae'])
# 训练数据做成字典，与输入层做对应
train_model_input = {name: train_data[name] for name in ['user_id', 'movie_id']}
history.fit(train_model_input,
train_data['label'].values,
batch_size=128,
epochs=2,
validation_split=0.2)
# 绘制网络结构图
plot_model(history,show_shapes=True)
