@(Aaron) [机器学习, EDA-数据探索性分析]
主要内容包括:
EDA的价值主要在于熟悉数据集,了解数据集,对数据集进行验证来确定所获得数据集可以用于接下来的机器学习或者深度学习使用。
当了解了数据集之后我们下一步就是要去了解变量间的相互关系以及变量与预测值之间的存在关系。
引导数据科学从业者进行数据处理以及特征工程的步骤,使数据集的结构和特征集让接下来的预测问题更加可靠。
完成对于数据的探索性分析,并对于数据进行一些图表或者文字总结并打卡。
以下库都是pip install 安装, 有特殊情况我会单独说明,例如 pip install pandas -i 常用库
#coding:utf-8
#导入warnings包,利用过滤器来实现忽略警告语句。
import warnings
warnings.filterwarnings('ignore')
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
import missingno as msno
## 1) 载入训练集和测试集;
path = 'dataset/'
Train_data = pd.read_csv(path+'used_car_train_20200313.csv', sep=' ')
Test_data = pd.read_csv(path+'used_car_testA_20200313.csv', sep=' ')
所有特征集均脱敏处理(方便大家观看)
## 2) 简略观察数据(head()+shape)
Train_data.head().append(Train_data.tail())
SaleID name regDate model brand bodyType fuelType gearbox power kilometer ... v_5 v_6 v_7 v_8 v_9 v_10 v_11 v_12 v_13 v_14
0 0 736 20040402 30.0 6 1.0 0.0 0.0 60 12.5 ... 0.235676 0.101988 0.129549 0.022816 0.097462 -2.881803 2.804097 -2.420821 0.795292 0.914762
1 1 2262 20030301 40.0 1 2.0 0.0 0.0 0 15.0 ... 0.264777 0.121004 0.135731 0.026597 0.020582 -4.900482 2.096338 -1.030483 -1.722674 0.245522
2 2 14874 20040403 115.0 15 1.0 0.0 0.0 163 12.5 ... 0.251410 0.114912 0.165147 0.062173 0.027075 -4.846749 1.803559 1.565330 -0.832687 -0.229963
3 3 71865 19960908 109.0 10 0.0 0.0 1.0 193 15.0 ... 0.274293 0.110300 0.121964 0.033395 0.000000 -4.509599 1.285940 -0.501868 -2.438353 -0.478699
4 4 111080 20120103 110.0 5 1.0 0.0 0.0 68 5.0 ... 0.228036 0.073205 0.091880 0.078819 0.121534 -1.896240 0.910783 0.931110 2.834518 1.923482
Test_data.shape
(50000, 30)
要养成看数据集的head()以及shape的习惯,这会让你每一步更放心,导致接下里的连串的错误, 如果对自己的pandas等操作不放心,建议执行一步看一下,这样会有效的方便你进行理解函数并进行操作
## 1) 通过describe()来熟悉数据的相关统计量
Train_data.describe()
SaleID name regDate model brand bodyType fuelType gearbox power kilometer ... v_5 v_6 v_7 v_8 v_9 v_10 v_11 v_12 v_13 v_14
count 150000.000000 150000.000000 1.500000e+05 149999.00 150000.00 145494.000 141320.000 144019.0000 150000.000 150000.000 ... 150000.0000 150000.000 150000.000 150000.000 150000.000 150000.000 150000.000 150000.000 150000.000 150000.000
mean 74999.500000 68349.172873 2.003417e+07 47.129021 8.052733 1.792369 0.375842 0.224943 119.316547 12.597160 ... 0.248204 0.044923 0.124692 0.058144 0.061996 -0.001000 0.009035 0.004813 0.000313 -0.000688
std 43301.414527 61103.875095 5.364988e+04 49.536040 7.864956 1.760640 0.548677 0.417546 177.168419 3.919576 ... 0.045804 0.051743 0.201410 0.029186 0.035692 3.772386 3.286071 2.517478 1.288988 1.038685
min 0.000000 0.000000 1.991000e+07 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000 0.500000 ... 0.000000 0.000000 0.000000 0.000000 0.000000 -9.168192 -5.558207 -9.639552 -4.153899 -6.546556
25% 37499.750000 11156.000000 1.999091e+07 10.000000 1.000000 0.000000 0.000000 0.000000 75.000000 12.500000 ... 0.243615 0.000038 0.062474 0.035334 0.033930 -3.722303 -1.951543 -1.871846 -1.057789 -0.437034
50% 74999.500000 51638.000000 2.003091e+07 30.000000 6.000000 1.000000 0.000000 0.000000 110.000000 15.000000 ... 0.257798 0.000812 0.095866 0.057014 0.058484 1.624076 -0.358053 -0.130753 -0.036245 0.141246
75% 112499.250000 118841.250000 2.007111e+07 66.000000 13.000000 3.000000 1.000000 0.000000 150.000000 15.000000 ... 0.265297 0.102009 0.125243 0.079382 0.087491 2.844357 1.255022 1.776933 0.942813 0.680378
max 149999.000000 196812.000000 2.015121e+07 247.000000 39.000000 7.000000 6.000000 1.000000 19312.0000 15.000000 ... 0.291838 0.151420 1.404936 0.160791 0.222787 12.357011 18.819042 13.847792 11.147669 8.658418
8 rows × 30 columns
Test_data.describe()
## 2) 通过info()来熟悉数据类型
Train_data.info()
RangeIndex: 150000 entries, 0 to 149999
Data columns (total 31 columns):
SaleID 150000 non-null int64
name 150000 non-null int64
regDate 150000 non-null int64
model 149999 non-null float64
brand 150000 non-null int64
bodyType 145494 non-null float64
fuelType 141320 non-null float64
gearbox 144019 non-null float64
power 150000 non-null int64
kilometer 150000 non-null float64
notRepairedDamage 150000 non-null object
regionCode 150000 non-null int64
seller 150000 non-null int64
offerType 150000 non-null int64
creatDate 150000 non-null int64
price 150000 non-null int64
v_0 150000 non-null float64
v_1 150000 non-null float64
v_2 150000 non-null float64
v_3 150000 non-null float64
v_4 150000 non-null float64
v_5 150000 non-null float64
v_6 150000 non-null float64
v_7 150000 non-null float64
v_8 150000 non-null float64
v_9 150000 non-null float64
v_10 150000 non-null float64
v_11 150000 non-null float64
v_12 150000 non-null float64
v_13 150000 non-null float64
v_14 150000 non-null float64
dtypes: float64(20), int64(10), object(1)
memory usage: 35.5+ MB
## 1) 查看每列的存在nan情况
Train_data.isnull().sum()
SaleID 0
name 0
regDate 0
model 1
brand 0
bodyType 4506
fuelType 8680
gearbox 5981
power 0
kilometer 0
notRepairedDamage 0
regionCode 0
seller 0
offerType 0
creatDate 0
price 0
v_0 0
v_1 0
v_2 0
v_3 0
v_4 0
v_5 0
v_6 0
v_7 0
v_8 0
v_9 0
v_10 0
v_11 0
v_12 0
v_13 0
v_14 0
dtype: int64
# nan可视化
missing = Train_data.isnull().sum()
missing = missing[missing > 0]
missing.sort_values(inplace=True)
missing.plot.bar()
通过以上两句可以很直观的了解哪些列存在 “nan”, 并可以把nan的个数打印,主要的目的在于 nan存在的个数是否真的很大,如果很小一般选择填充,如果使用lgb等树模型可以直接空缺,让树自己去优化,但如果nan存在的过多、可以考虑删掉
# 可视化看下缺省值
msno.matrix(Train_data.sample(250))
msno.bar(Train_data.sample(1000))
# 可视化看下缺省值
msno.matrix(Test_data.sample(250))
msno.bar(Test_data.sample(1000))
测试集的缺省和训练集的差不多情况, 可视化有四列有缺省,notRepairedDamage缺省得最多
## 2) 查看异常值检测
Train_data.info()
RangeIndex: 150000 entries, 0 to 149999
Data columns (total 31 columns):
SaleID 150000 non-null int64
name 150000 non-null int64
regDate 150000 non-null int64
model 149999 non-null float64
brand 150000 non-null int64
bodyType 145494 non-null float64
fuelType 141320 non-null float64
gearbox 144019 non-null float64
power 150000 non-null int64
kilometer 150000 non-null float64
notRepairedDamage 150000 non-null object
regionCode 150000 non-null int64
seller 150000 non-null int64
offerType 150000 non-null int64
creatDate 150000 non-null int64
price 150000 non-null int64
v_0 150000 non-null float64
v_1 150000 non-null float64
v_2 150000 non-null float64
v_3 150000 non-null float64
v_4 150000 non-null float64
v_5 150000 non-null float64
v_6 150000 non-null float64
v_7 150000 non-null float64
v_8 150000 non-null float64
v_9 150000 non-null float64
v_10 150000 non-null float64
v_11 150000 non-null float64
v_12 150000 non-null float64
v_13 150000 non-null float64
v_14 150000 non-null float64
dtypes: float64(20), int64(10), object(1)
memory usage: 35.5+ MB
可以发现除了notRepairedDamage 为object类型其他都为数字 这里我们把他的几个不同的值都进行显示就知道了
Train_data['notRepairedDamage'].value_counts()
0.0 111361
'-' 24324
1.0 14315
Name: notRepairedDamage, dtype: int64
可以看出来‘ - ’也为空缺值,因为很多模型对nan有直接的处理,这里我们先不做处理,先替换成nan
Train_data['notRepairedDamage'].replace('-', np.nan, inplace=True)
Train_data['notRepairedDamage'].value_counts()
## 2) 查看异常值检测
Train_data.info()
0.0 111361
1.0 14315
Name: notRepairedDamage, dtype: int64
Train_data.isnull().sum()
SaleID 0
name 0
regDate 0
model 1
brand 0
bodyType 4506
fuelType 8680
gearbox 5981
power 0
kilometer 0
notRepairedDamage 24324
regionCode 0
seller 0
offerType 0
creatDate 0
price 0
v_0 0
v_1 0
v_2 0
v_3 0
v_4 0
v_5 0
v_6 0
v_7 0
v_8 0
v_9 0
v_10 0
v_11 0
v_12 0
v_13 0
v_14 0
dtype: int64
Test_data['notRepairedDamage'].value_counts()
0.0 37249
- 8031
1.0 4720
Name: notRepairedDamage, dtype: int64
Test_data['notRepairedDamage'].replace('-', np.nan, inplace=True)
以下两个类别特征严重倾斜,一般不会对预测有什么帮助,故这边先删掉,当然你也可以继续挖掘,但是一般意义不大
Train_data["seller"].value_counts()
0 149999
1 1
Name: seller, dtype: int64
Train_data["offerType"].value_counts()
## 2) 查看异常值检测
Train_data.info()
0 150000
Name: offerType, dtype: int64
del Train_data["seller"]
del Train_data["offerType"]
del Test_data["seller"]
del Test_data["offerType"]
Train_data['price']
Train_data['price'].value_counts()
## 1) 总体分布概况(无界约翰逊分布等)
import scipy.stats as st
y = Train_data['price']
plt.figure(1); plt.title('Johnson SU')
sns.distplot(y, kde=False, fit=st.johnsonsu)
plt.figure(2); plt.title('Normal')
sns.distplot(y, kde=False, fit=st.norm)
plt.figure(3); plt.title('Log Normal')
sns.distplot(y, kde=False, fit=st.lognorm)
价格不服从正态分布,所以在进行回归之前,它必须进行转换。虽然对数变换做得很好,但最佳拟合是无界约翰逊分布
## 2) 查看skewness and kurtosis
sns.distplot(Train_data['price']);
print("Skewness: %f" % Train_data['price'].skew())
print("Kurtosis: %f" % Train_data['price'].kurt())
Skewness: 3.346487
Kurtosis: 18.995183
Train_data.skew(), Train_data.kurt()
sns.distplot(Train_data.skew(),color='blue',axlabel ='Skewness')
sns.distplot(Train_data.kurt(),color='orange',axlabel ='Kurtness')
skew、kurt说明参考https://www.cnblogs.com/wyy1480/p/10474046.html
## 3) 查看预测值的具体频数
plt.hist(Train_data['price'], orientation = 'vertical',histtype = 'bar', color ='red')
plt.show()
查看频数, 大于20000得值极少,其实这里也可以把这些当作特殊得值(异常值)直接用填充或者删掉,再前面进行
# log变换 z之后的分布较均匀,可以进行log变换进行预测,这也是预测问题常用的trick
plt.hist(np.log(Train_data['price']), orientation = 'vertical',histtype = 'bar', color ='red')
plt.show()
数据类型
# 分离label即预测值
Y_train = Train_data['price']
# 这个区别方式适用于没有直接label coding的数据
# 这里不适用,需要人为根据实际含义来区分
# 数字特征
# numeric_features = Train_data.select_dtypes(include=[np.number])
# numeric_features.columns
# # 类型特征
# categorical_features = Train_data.select_dtypes(include=[np.object])
# categorical_features.columns
numeric_features = ['power', 'kilometer', 'v_0', 'v_1', 'v_2', 'v_3', 'v_4', 'v_5', 'v_6', 'v_7', 'v_8', 'v_9', 'v_10', 'v_11', 'v_12', 'v_13','v_14' ]
categorical_features = ['name', 'model', 'brand', 'bodyType', 'fuelType', 'gearbox', 'notRepairedDamage', 'regionCode',]
# 特征nunique分布
for cat_fea in categorical_features:
print(cat_fea + "的特征分布如下:")
print("{}特征有个{}不同的值".format(cat_fea, Train_data[cat_fea].nunique()))
print(Train_data[cat_fea].value_counts())
# 特征nunique分布
for cat_fea in categorical_features:
print(cat_fea + "的特征分布如下:")
print("{}特征有个{}不同的值".format(cat_fea, Test_data[cat_fea].nunique()))
print(Test_data[cat_fea].value_counts())
numeric_features.append('price')
## 1) 相关性分析
price_numeric = Train_data[numeric_features]
correlation = price_numeric.corr()
print(correlation['price'].sort_values(ascending = False),'\n')
price 1.000000
v_12 0.692823
v_8 0.685798
v_0 0.628397
power 0.219834
v_5 0.164317
v_2 0.085322
v_6 0.068970
v_1 0.060914
v_14 0.035911
v_13 -0.013993
v_7 -0.053024
v_4 -0.147085
v_9 -0.206205
v_10 -0.246175
v_11 -0.275320
kilometer -0.440519
v_3 -0.730946
Name: price, dtype: float64
f , ax = plt.subplots(figsize = (7, 7))
plt.title('Correlation of Numeric Features with Price',y=1,size=16)
sns.heatmap(correlation,square = True, vmax=0.8)
del price_numeric['price']
## 2) 查看几个特征得 偏度和峰值
for col in numeric_features:
print('{:15}'.format(col),
'Skewness: {:05.2f}'.format(Train_data[col].skew()) ,
' ' ,
'Kurtosis: {:06.2f}'.format(Train_data[col].kurt())
)
power Skewness: 65.86 Kurtosis: 5733.45
kilometer Skewness: -1.53 Kurtosis: 001.14
v_0 Skewness: -1.32 Kurtosis: 003.99
v_1 Skewness: 00.36 Kurtosis: -01.75
v_2 Skewness: 04.84 Kurtosis: 023.86
v_3 Skewness: 00.11 Kurtosis: -00.42
v_4 Skewness: 00.37 Kurtosis: -00.20
v_5 Skewness: -4.74 Kurtosis: 022.93
v_6 Skewness: 00.37 Kurtosis: -01.74
v_7 Skewness: 05.13 Kurtosis: 025.85
v_8 Skewness: 00.20 Kurtosis: -00.64
v_9 Skewness: 00.42 Kurtosis: -00.32
v_10 Skewness: 00.03 Kurtosis: -00.58
v_11 Skewness: 03.03 Kurtosis: 012.57
v_12 Skewness: 00.37 Kurtosis: 000.27
v_13 Skewness: 00.27 Kurtosis: -00.44
v_14 Skewness: -1.19 Kurtosis: 002.39
price Skewness: 03.35 Kurtosis: 019.00
## 3) 每个数字特征得分布可视化
f = pd.melt(Train_data, value_vars=numeric_features)
g = sns.FacetGrid(f, col="variable", col_wrap=2, sharex=False, sharey=False)
g = g.map(sns.distplot, "value")
## 4) 数字特征相互之间的关系可视化
sns.set()
columns = ['price', 'v_12', 'v_8' , 'v_0', 'power', 'v_5', 'v_2', 'v_6', 'v_1', 'v_14']
sns.pairplot(Train_data[columns],size = 2 ,kind ='scatter',diag_kind='kde')
plt.show()
此处是多变量之间的关系可视化,可视化更多学习可参考很不错的文章 https://www.jianshu.com/p/6e18d21a4cad
## 5) 多变量互相回归关系可视化
fig, ((ax1, ax2), (ax3, ax4), (ax5, ax6), (ax7, ax8), (ax9, ax10)) = plt.subplots(nrows=5, ncols=2, figsize=(24, 20))
# ['v_12', 'v_8' , 'v_0', 'power', 'v_5', 'v_2', 'v_6', 'v_1', 'v_14']
v_12_scatter_plot = pd.concat([Y_train,Train_data['v_12']],axis = 1)
sns.regplot(x='v_12',y = 'price', data = v_12_scatter_plot,scatter= True, fit_reg=True, ax=ax1)
v_8_scatter_plot = pd.concat([Y_train,Train_data['v_8']],axis = 1)
sns.regplot(x='v_8',y = 'price',data = v_8_scatter_plot,scatter= True, fit_reg=True, ax=ax2)
v_0_scatter_plot = pd.concat([Y_train,Train_data['v_0']],axis = 1)
sns.regplot(x='v_0',y = 'price',data = v_0_scatter_plot,scatter= True, fit_reg=True, ax=ax3)
power_scatter_plot = pd.concat([Y_train,Train_data['power']],axis = 1)
sns.regplot(x='power',y = 'price',data = power_scatter_plot,scatter= True, fit_reg=True, ax=ax4)
v_5_scatter_plot = pd.concat([Y_train,Train_data['v_5']],axis = 1)
sns.regplot(x='v_5',y = 'price',data = v_5_scatter_plot,scatter= True, fit_reg=True, ax=ax5)
v_2_scatter_plot = pd.concat([Y_train,Train_data['v_2']],axis = 1)
sns.regplot(x='v_2',y = 'price',data = v_2_scatter_plot,scatter= True, fit_reg=True, ax=ax6)
v_6_scatter_plot = pd.concat([Y_train,Train_data['v_6']],axis = 1)
sns.regplot(x='v_6',y = 'price',data = v_6_scatter_plot,scatter= True, fit_reg=True, ax=ax7)
v_1_scatter_plot = pd.concat([Y_train,Train_data['v_1']],axis = 1)
sns.regplot(x='v_1',y = 'price',data = v_1_scatter_plot,scatter= True, fit_reg=True, ax=ax8)
v_14_scatter_plot = pd.concat([Y_train,Train_data['v_14']],axis = 1)
sns.regplot(x='v_14',y = 'price',data = v_14_scatter_plot,scatter= True, fit_reg=True, ax=ax9)
v_13_scatter_plot = pd.concat([Y_train,Train_data['v_13']],axis = 1)
sns.regplot(x='v_13',y = 'price',data = v_13_scatter_plot,scatter= True, fit_reg=True, ax=ax10)
## 1) unique分布
for fea in categorical_features:
print(Train_data[fea].nunique())
print(categorical_features)
## 2) 类别特征箱形图可视化
# 因为 name和 regionCode的类别太稀疏了,这里我们把不稀疏的几类画一下
categorical_features = ['model',
'brand',
'bodyType',
'fuelType',
'gearbox',
'notRepairedDamage']
for c in categorical_features:
Train_data[c] = Train_data[c].astype('category')
if Train_data[c].isnull().any():
Train_data[c] = Train_data[c].cat.add_categories(['MISSING'])
Train_data[c] = Train_data[c].fillna('MISSING')
def boxplot(x, y, **kwargs):
sns.boxplot(x=x, y=y)
x=plt.xticks(rotation=90)
f = pd.melt(Train_data, id_vars=['price'], value_vars=categorical_features)
g = sns.FacetGrid(f, col="variable", col_wrap=2, sharex=False, sharey=False, size=5)
g = g.map(boxplot, "value", "price")
print(Train_data.columns)
## 3) 类别特征的小提琴图可视化
catg_list = categorical_features
target = 'price'
for catg in catg_list :
sns.violinplot(x=catg, y=target, data=Train_data)
plt.show()
categorical_features = ['model',
'brand',
'bodyType',
'fuelType',
'gearbox',
'notRepairedDamage']
## 4) 类别特征的柱形图可视化
def bar_plot(x, y, **kwargs):
sns.barplot(x=x, y=y)
x=plt.xticks(rotation=90)
f = pd.melt(Train_data, id_vars=['price'], value_vars=categorical_features)
g = sns.FacetGrid(f, col="variable", col_wrap=2, sharex=False, sharey=False, size=5)
g = g.map(bar_plot, "value", "price")
## 5) 类别特征的每个类别频数可视化(count_plot)
def count_plot(x, **kwargs):
sns.countplot(x=x)
x=plt.xticks(rotation=90)
f = pd.melt(Train_data, value_vars=categorical_features)
g = sns.FacetGrid(f, col="variable", col_wrap=2, sharex=False, sharey=False, size=5)
g = g.map(count_plot, "value")
用pandas_profiling生成一个较为全面的可视化和数据报告(较为简单、方便) 最终打开html文件即可
import pandas_profiling
pfr = pandas_profiling.ProfileReport(Train_data)
pfr.to_file("./example.html")
所给出的EDA步骤为广为普遍的步骤,在实际的不管是工程还是比赛过程中,这只是最开始的一步,也是最基本的一步。
接下来一般要结合模型的效果以及特征工程等来分析数据的实际建模情况,根据自己的一些理解,查阅文献,对实际问题做出判断和深入的理解。
最后不断进行EDA与数据处理和挖掘,来到达更好的数据结构和分布以及较为强势相关的特征
数据探索在机器学习中我们一般称为EDA(Exploratory Data Analysis):
是指对已有的数据(特别是调查或观察得来的原始数据)在尽量少的先验假定下进行探索,通过作图、制表、方程拟合、计算特征量等手段探索数据的结构和规律的一种数据分析方法。
数据探索有利于我们发现数据的一些特性,数据之间的关联性,对于后续的特征构建是很有帮助的。
对于数据的初步分析(直接查看数据,或.sum(), .mean(),.descirbe()等统计函数)可以从:样本数量,训练集数量,是否有时间特征,是否是时许问题,特征所表示的含义(非匿名特征),特征类型(字符类似,int,float,time),特征的缺失情况(注意缺失的在数据中的表现形式,有些是空的有些是”NAN”符号等),特征的均值方差情况。
分析记录某些特征值缺失占比30%以上样本的缺失处理,有助于后续的模型验证和调节,分析特征应该是填充(填充方式是什么,均值填充,0填充,众数填充等),还是舍去,还是先做样本分类用不同的特征模型去预测。
对于异常值做专门的分析,分析特征异常的label是否为异常值(或者偏离均值较远或者事特殊符号),异常值是否应该剔除,还是用正常值填充,是记录异常,还是机器本身异常等。
对于Label做专门的分析,分析标签的分布情况等。
进步分析可以通过对特征作图,特征和label联合做图(统计图,离散图),直观了解特征的分布情况,通过这一步也可以发现数据之中的一些异常值等,通过箱型图分析一些特征值的偏离情况,对于特征和特征联合作图,对于特征和label联合作图,分析其中的一些关联性。