import numpy as np
import pandas as pd
from tensorflow.keras.layers import *
import tensorflow.keras.backend as K
import matplotlib.pyplot as plt
import tensorflow as tf
from tensorflow.keras.models import Model
from tensorflow.keras.utils import plot_model
from sklearn.preprocessing import LabelEncoder
def dataset(path):
'''数据获取以及初步处理'''
data = pd.read_csv(path)
cols = data.columns.values # 获取数据的标签，方便之后进行数据处理
dense_feats = [i for i in cols if i[0] == "I"] # 数据主要分为两种类型
sparse_feats = [j for j in cols if j[0] == "C"]
#首先对标签为dense_feats的数据进行处理
data_dense = data.copy() # 保留原数据
data_dense = data_dense[dense_feats].fillna(0.0) # 缺失值填补
for f in dense_feats:
data_dense[f] = data_dense[f].apply(lambda x: np.log(x+1) if x > -1 else -1)
# 再对标签为sparse_feats的数据进行处理
data_sparse = data.copy()
data_sparse = data_sparse[sparse_feats].fillna("-1") # 空白值补-1
for f in sparse_feats:
label_encoder = LabelEncoder() # 将数据离散化
data_sparse[f] = label_encoder.fit_transform(data_sparse[f])
total_data = pd.concat([data_dense, data_sparse], axis=1) #再将两种数据联立起来
total_data['label'] = data['label']
return dense_feats, sparse_feats, data_dense, data_sparse, total_data
def data_deal(dense_feats, sparse_feats, data_dense, data_sparse):
'''对数据进行特征处理'''
# 先构造 dense 特征的输入
dense_inputs = []
for f in dense_feats:
_input = Input([1], name=f)
dense_inputs.append(_input)
# 将输入拼接到一起，方便连接 Dense 层
# 若axis=0，则要求除了a.shape[0]和b.shape[0]可以不等之外，其它维度必须相等。此时c.shape[0] = a.shape[0]+b.shape[0]
# 若axis=1，则要求除了a.shape[1]和b.shape[1]可以不等之外，其它维度必须相等。此时c.shape[1] = a.shape[1]+b.shape[1]
concat_dense_inputs = Concatenate(axis=1)(dense_inputs)
# 然后连上输出为1个单元的全连接层，表示对 dense 变量的加权求和
fst_order_dense_layer = Dense(1)(concat_dense_inputs)
# 再对 sparse 特征构造输入，目的是方便后面构造二阶组合特征
sparse_inputs = []
for f in sparse_feats:
_input = Input([1], name=f)
sparse_inputs.append(_input)
sparse_1d_embed = []
for i, _input in enumerate(sparse_inputs):
f = sparse_feats[i]
voc_size = total_data[f].nunique()
# 使用 l2 正则化防止过拟合
reg = tf.keras.regularizers.l2(0.5)
_embed = Embedding(voc_size, 1, embeddings_regularizer=reg)(_input)
# 由于 Embedding 的结果是二维的，
# 因此如果需要在 Embedding 之后加入 Dense 层，则需要先连接上 Flatten 层
_embed = Flatten()(_embed)
sparse_1d_embed.append(_embed)
# 对每个 embedding lookup 的结果 wi 求和
fst_order_sparse_layer = Add()(sparse_1d_embed)
# 再将dense 和 sparse的特征联合在一起
linear_part = Add()([fst_order_dense_layer, fst_order_sparse_layer])
return linear_part,dense_inputs,sparse_inputs
def embedding(sparse_inputs, sparse_feats, total_data):
'''进行特征组合之前，每个 sparse 特征需要先进行 embedding'''
k = 8
# 只考虑sparse的二阶交叉
sparse_kd_embed = []
for i, _input in enumerate(sparse_inputs):
f = sparse_feats[i]
voc_size = total_data[f].nunique()
reg = tf.keras.regularizers.l2(0.7) # 防止过拟合
# embedding为嵌入层，嵌入层可以降维，也可以升维，这里是降维
_embed = Embedding(voc_size, k, embeddings_regularizer=reg)(_input)
sparse_kd_embed.append(_embed)
#接下来就是要进行特征组合，如果对 n 个 sparse 特征两两组合，那么复杂度应该是O（n^），但是可以用计算技巧降为O(kn)
# 1.将所有sparse的embedding拼接起来，得到 (n, k)的矩阵，其中n为特征数，k为embedding大小
concat_sparse_kd_embed = Concatenate(axis=1)(sparse_kd_embed)
# 2.先求和再平方
sum_kd_embed = Lambda(lambda x: K.sum(x, axis=1))(concat_sparse_kd_embed)
square_sum_kd_embed = Multiply()([sum_kd_embed, sum_kd_embed])
# 3.先平方再求和
square_kd_embed = Multiply()([concat_sparse_kd_embed, concat_sparse_kd_embed])
sum_square_kd_embed = Lambda(lambda x: K.sum(x, axis=1))(square_kd_embed)
# 4.相减除以2
sub = Subtract()([square_sum_kd_embed, sum_square_kd_embed])
sub = Lambda(lambda x: x*0.5)(sub)
snd_order_sparse_layer = Lambda(lambda x: K.sum(x, axis=1, keepdims=True))(sub)
return snd_order_sparse_layer, concat_sparse_kd_embed
def DNN(concat_sparse_kd_embed,linear_part, snd_order_sparse_layer):
'''接下来构造DNN部分'''
flatten_sparse_embed = Flatten()(concat_sparse_kd_embed)
fc_layer = Dropout(0.5)(Dense(256, activation='relu')(flatten_sparse_embed))
fc_layer = Dropout(0.3)(Dense(256, activation='relu')(fc_layer))
fc_layer = Dropout(0.1)(Dense(256, activation='relu')(fc_layer))
fc_layer_output = Dense(1)(fc_layer)
output_layer = Add()([linear_part, snd_order_sparse_layer, fc_layer_output])
output_layer = Activation("sigmoid")(output_layer)
return output_layer, fc_layer_output
def add_layer(linear_part, snd_order_sparse_layer, fc_layer_output):
#将dnn和fm联立起来
output_layer = Add()([linear_part, snd_order_sparse_layer, fc_layer_output])
output_layer = Activation("sigmoid")(output_layer)
model = Model(dense_inputs+sparse_inputs, output_layer)
model.compile(optimizer="adam",
loss="binary_crossentropy",
metrics=["binary_crossentropy", tf.keras.metrics.AUC(name='auc')])
return model
def train(total_data,dense_feats,sparse_feats,model):
train_data = total_data.loc[:500000-1]
valid_data = total_data.loc[500000:]
train_dense_x = [train_data[f].values for f in dense_feats]
train_sparse_x = [train_data[f].values for f in sparse_feats]
train_label = [train_data['label'].values]
val_dense_x = [valid_data[f].values for f in dense_feats]
val_sparse_x = [valid_data[f].values for f in sparse_feats]
val_label = [valid_data['label'].values]
model.fit(train_dense_x+train_sparse_x,
train_label, epochs=7, batch_size=256,
validation_data=(val_dense_x+val_sparse_x, val_label),
)
model.save('my_mode.h5')
if __name__ == '__main__':
dense_feats, sparse_feats, data_dense, data_sparse, total_data = dataset('criteo_sampled_data.csv')
linear_part,dense_inputs,sparse_inputs = data_deal(dense_feats, sparse_feats, data_dense, data_sparse)
snd_order_sparse_layer, concat_sparse_kd_embed = embedding(sparse_inputs, sparse_feats, total_data)
output_layer, fc_layer_output = DNN(concat_sparse_kd_embed,linear_part, snd_order_sparse_layer)
model = add_layer(linear_part, snd_order_sparse_layer, fc_layer_output)
train(total_data,dense_feats,sparse_feats,model)
