2021-03-21 : task03:DeeFM模型学习总结

1. 简述了模型部分的各个操作
2. 对代码进行了结果展示
3. 向量的特征交叉测试(FM部分)

1.模型部分

1. 数值型特征的输入层
2. 类别型特征的输入层
3. 数值型特征不用Embedding
4. 类别型特征的Embedding
   • 与liner部分结合 - Embedding为1维->flaten->Add->与Dense后的数值型Add()
   • FM部分:类别型特征的特征交叉 - 不需要Flaten
   • deep: Embedding为多维(4维),然后特征交叉
5. 一阶特征相加
   • 数值型特征的concat ->Dense + 与类别型特征的add后的add
6. fm二阶交叉
   • 文章最后自己测试了下特征交叉(就是(a2-b2)/2) – 主要是掌握
   • 个人认为可以用layers.Dot()层实现特征交叉,水平有限,有待尝试
7. deep部分
   • 多维EmbeddingFlatten后->拼接concat->多层神经网络迭代
8. 三合一
   • Add()

2.代码展示

# 导包
import warnings
warnings.filterwarnings("ignore")
import pandas as pd
import numpy as np
import tensorflow as tf
from tensorflow.keras.layers import *
from tensorflow.keras.models import *
from sklearn.preprocessing import  MinMaxScaler, LabelEncoder
# from utils import SparseFeat, DenseFeat, VarLenSparseFeat

# 读取数据
data = pd.read_csv('./data/criteo_sample.txt')

data.head()

	label	I1	I2	I3	I4	I5	I6	I7	I8	I9	...	C17	C18	C19	C20	C21	C22	C23	C24	C25	C26
0	0	NaN	3	260.0	NaN	17668.0	NaN	NaN	33.0	NaN	...	e5ba7672	87c6f83c	NaN	NaN	0429f84b	NaN	3a171ecb	c0d61a5c	NaN	NaN
1	0	NaN	-1	19.0	35.0	30251.0	247.0	1.0	35.0	160.0	...	d4bb7bd8	6fc84bfb	NaN	NaN	5155d8a3	NaN	be7c41b4	ded4aac9	NaN	NaN
2	0	0.0	0	2.0	12.0	2013.0	164.0	6.0	35.0	523.0	...	e5ba7672	675c9258	NaN	NaN	2e01979f	NaN	bcdee96c	6d5d1302	NaN	NaN
3	0	NaN	13	1.0	4.0	16836.0	200.0	5.0	4.0	29.0	...	e5ba7672	52e44668	NaN	NaN	e587c466	NaN	32c7478e	3b183c5c	NaN	NaN
4	0	0.0	0	104.0	27.0	1990.0	142.0	4.0	32.0	37.0	...	e5ba7672	25c88e42	21ddcdc9	b1252a9d	0e8585d2	NaN	32c7478e	0d4a6d1a	001f3601	92c878de

5 rows × 40 columns

# 划分出类别型特征和数值型特征
columns = data.columns.values
dense_features = [f for f in columns if 'I' in f]
sparse_features = [f for f in columns if 'C' in f]

columns

array(['label', 'I1', 'I2', 'I3', 'I4', 'I5', 'I6', 'I7', 'I8', 'I9',
       'I10', 'I11', 'I12', 'I13', 'C1', 'C2', 'C3', 'C4', 'C5', 'C6',
       'C7', 'C8', 'C9', 'C10', 'C11', 'C12', 'C13', 'C14', 'C15', 'C16',
       'C17', 'C18', 'C19', 'C20', 'C21', 'C22', 'C23', 'C24', 'C25',
       'C26'], dtype=object)

dense_features

['I1',
 'I2',
 'I3',
 'I4',
 'I5',
 'I6',
 'I7',
 'I8',
 'I9',
 'I10',
 'I11',
 'I12',
 'I13']

# 数据处理 : 分别处理数值型特征和类别型特征
def data_process(data_df,dense_festures,sparse_features):
    """
    对特征进行处理
    """
    # 处理数值型特征
    data_df[dense_festures]= data_df[dense_features].fillna(0.0) # 所有的数值列的空值补0
    for f in dense_features: # 每一列数值型特征的归一化
        data_df[f] = data_df[f].apply(lambda x:np.log(x+1) if x>-1 else -1 )
    # 处理类别型特征
    data_df[sparse_features] = data_df[sparse_features].fillna('-1')
    for f in sparse_features:
        # 对每列的类别型特征进行数值化处理
        le = LabelEncoder()
        data_df[f] = le.fit_transform(data_df[f])
    return data_df[dense_features+sparse_features]

train_data = data_process(data,dense_features,sparse_features)

train_data

	I1	I2	I3	I4	I5	I6	I7	I8	I9	I10	...	C17	C18	C19	C20	C21	C22	C23	C24	C25	C26
0	0.000000	1.386294	5.564520	0.000000	9.779567	0.000000	0.000000	3.526361	0.000000	0.000000	...	8	66	0	0	3	0	1	96	0	0
1	0.000000	-1.000000	2.995732	3.583519	10.317318	5.513429	0.693147	3.583519	5.081404	0.000000	...	7	52	0	0	47	0	7	112	0	0
2	0.000000	0.000000	1.098612	2.564949	7.607878	5.105945	1.945910	3.583519	6.261492	0.000000	...	8	49	0	0	25	0	6	53	0	0
3	0.000000	2.639057	0.693147	1.609438	9.731334	5.303305	1.791759	1.609438	3.401197	0.000000	...	8	37	0	0	156	0	0	32	0	0
4	0.000000	0.000000	4.653960	3.332205	7.596392	4.962845	1.609438	3.496508	3.637586	0.000000	...	8	14	5	3	9	0	0	5	1	47
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
195	0.000000	0.000000	4.736198	1.386294	8.018625	6.356108	1.098612	1.386294	5.370638	0.000000	...	0	74	5	1	30	5	0	118	17	48
196	0.000000	0.693147	0.693147	0.693147	7.382746	2.564949	0.693147	2.564949	2.772589	0.000000	...	1	25	0	0	138	0	0	68	0	0
197	0.693147	0.000000	1.945910	1.386294	0.000000	0.000000	2.995732	1.386294	1.386294	0.693147	...	4	40	17	2	41	0	0	12	16	11
198	0.000000	3.135494	1.945910	3.135494	5.318120	5.036953	4.394449	2.944439	6.232448	0.000000	...	4	7	18	1	123	0	0	10	16	49
199	0.693147	-1.000000	0.000000	0.000000	4.934474	0.000000	0.693147	0.000000	0.000000	0.693147	...	7	72	0	0	0	0	0	0	0	0

200 rows × 39 columns

train_data['label'] = data['label']

train_data

	I1	I2	I3	I4	I5	I6	I7	I8	I9	I10	...	C18	C19	C20	C21	C22	C23	C24	C25	C26	label
0	0.000000	1.386294	5.564520	0.000000	9.779567	0.000000	0.000000	3.526361	0.000000	0.000000	...	66	0	0	3	0	1	96	0	0	0
1	0.000000	-1.000000	2.995732	3.583519	10.317318	5.513429	0.693147	3.583519	5.081404	0.000000	...	52	0	0	47	0	7	112	0	0	0
2	0.000000	0.000000	1.098612	2.564949	7.607878	5.105945	1.945910	3.583519	6.261492	0.000000	...	49	0	0	25	0	6	53	0	0	0
3	0.000000	2.639057	0.693147	1.609438	9.731334	5.303305	1.791759	1.609438	3.401197	0.000000	...	37	0	0	156	0	0	32	0	0	0
4	0.000000	0.000000	4.653960	3.332205	7.596392	4.962845	1.609438	3.496508	3.637586	0.000000	...	14	5	3	9	0	0	5	1	47	0
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
195	0.000000	0.000000	4.736198	1.386294	8.018625	6.356108	1.098612	1.386294	5.370638	0.000000	...	74	5	1	30	5	0	118	17	48	0
196	0.000000	0.693147	0.693147	0.693147	7.382746	2.564949	0.693147	2.564949	2.772589	0.000000	...	25	0	0	138	0	0	68	0	0	1
197	0.693147	0.000000	1.945910	1.386294	0.000000	0.000000	2.995732	1.386294	1.386294	0.693147	...	40	17	2	41	0	0	12	16	11	1
198	0.000000	3.135494	1.945910	3.135494	5.318120	5.036953	4.394449	2.944439	6.232448	0.000000	...	7	18	1	123	0	0	10	16	49	0
199	0.693147	-1.000000	0.000000	0.000000	4.934474	0.000000	0.693147	0.000000	0.000000	0.693147	...	72	0	0	0	0	0	0	0	0	0

200 rows × 40 columns

train_data.info()

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 200 entries, 0 to 199
Data columns (total 40 columns):
 #   Column  Non-Null Count  Dtype  
---  ------  --------------  -----  
 0   I1      200 non-null    float64
 1   I2      200 non-null    float64
 2   I3      200 non-null    float64
 3   I4      200 non-null    float64
 4   I5      200 non-null    float64
 5   I6      200 non-null    float64
 6   I7      200 non-null    float64
 7   I8      200 non-null    float64
 8   I9      200 non-null    float64
 9   I10     200 non-null    float64
 10  I11     200 non-null    float64
 11  I12     200 non-null    float64
 12  I13     200 non-null    float64
 13  C1      200 non-null    int32  
 14  C2      200 non-null    int32  
 15  C3      200 non-null    int32  
 16  C4      200 non-null    int32  
 17  C5      200 non-null    int32  
 18  C6      200 non-null    int32  
 19  C7      200 non-null    int32  
 20  C8      200 non-null    int32  
 21  C9      200 non-null    int32  
 22  C10     200 non-null    int32  
 23  C11     200 non-null    int32  
 24  C12     200 non-null    int32  
 25  C13     200 non-null    int32  
 26  C14     200 non-null    int32  
 27  C15     200 non-null    int32  
 28  C16     200 non-null    int32  
 29  C17     200 non-null    int32  
 30  C18     200 non-null    int32  
 31  C19     200 non-null    int32  
 32  C20     200 non-null    int32  
 33  C21     200 non-null    int32  
 34  C22     200 non-null    int32  
 35  C23     200 non-null    int32  
 36  C24     200 non-null    int32  
 37  C25     200 non-null    int32  
 38  C26     200 non-null    int32  
 39  label   200 non-null    int64  
dtypes: float64(13), int32(26), int64(1)
memory usage: 42.3 KB

模型的创建

# 1. 数值型特征的输入层
dense_inputs=[]
for f in dense_features:
    _input = Input([1],name=f)
    dense_inputs.append(_input)

dense_inputs

[<KerasTensor: shape=(None, 1) dtype=float32 (created by layer 'I1')>,
 <KerasTensor: shape=(None, 1) dtype=float32 (created by layer 'I2')>,
 <KerasTensor: shape=(None, 1) dtype=float32 (created by layer 'I3')>,
 <KerasTensor: shape=(None, 1) dtype=float32 (created by layer 'I4')>,
 <KerasTensor: shape=(None, 1) dtype=float32 (created by layer 'I5')>,
 <KerasTensor: shape=(None, 1) dtype=float32 (created by layer 'I6')>,
 <KerasTensor: shape=(None, 1) dtype=float32 (created by layer 'I7')>,
 <KerasTensor: shape=(None, 1) dtype=float32 (created by layer 'I8')>,
 <KerasTensor: shape=(None, 1) dtype=float32 (created by layer 'I9')>,
 <KerasTensor: shape=(None, 1) dtype=float32 (created by layer 'I10')>,
 <KerasTensor: shape=(None, 1) dtype=float32 (created by layer 'I11')>,
 <KerasTensor: shape=(None, 1) dtype=float32 (created by layer 'I12')>,
 <KerasTensor: shape=(None, 1) dtype=float32 (created by layer 'I13')>]

# 2. 类别型特征的输入层
sparse_inputs=[]
for f in sparse_features:
    _input = Input([