预测信用卡流失客户

kaggle数据地址：https://www.kaggle.com/sakshigoyal7/credit-card-customers

导入数据

#导入模块
import pandas as pd 
import numpy as np 
import  matplotlib.pyplot as plt
import seaborn as sns

BankChurners = pd.read_csv(\'D:\\python_home\\预测客户流失\\bankchurners\\BankChurners.csv\')

 

简单的数据查看

BankChurners.columns

\'\'\'
Index([\'CLIENTNUM\', \'Attrition_Flag\', \'Customer_Age\', \'Gender\',
       \'Dependent_count\', \'Education_Level\', \'Marital_Status\',
       \'Income_Category\', \'Card_Category\', \'Months_on_book\',
       \'Total_Relationship_Count\', \'Months_Inactive_12_mon\',
       \'Contacts_Count_12_mon\', \'Credit_Limit\', \'Total_Revolving_Bal\',
       \'Avg_Open_To_Buy\', \'Total_Amt_Chng_Q4_Q1\', \'Total_Trans_Amt\',
       \'Total_Trans_Ct\', \'Total_Ct_Chng_Q4_Q1\', \'Avg_Utilization_Ratio\',
       \'Naive_Bayes_Classifier_Attrition_Flag_Card_Category_Contacts_Count_12_mon_Dependent_count_Education_Level_Months_Inactive_12_mon_1\',
       \'Naive_Bayes_Classifier_Attrition_Flag_Card_Category_Contacts_Count_12_mon_Dependent_count_Education_Level_Months_Inactive_12_mon_2\'],
      dtype=\'object\')
\'\'\'

拿到百度翻译了一下

\'\'\'
\'CLIENTNUM\'，\'消耗标志\'，\'客户年龄\'，\'性别\'，

“受抚养人数量”、“受教育程度”、“婚姻状况”，

\'收入类别\'，\'卡片类别\'，\'账面上的月份\'，

\'总关系数\'，\'月数\'u不活跃\'u 12个月\'，

“联系人数量12个月”，“信用额度”，“总周转余额”，

\'平均开盘价到买入价\'，\'总金额\'，

\'总交易量\'，\'第4季度总交易量\'，\'平均利用率\'，

\'天真的_Bayes_分类器_消耗_标志_卡片_类别_联系人_Count_12 _mon_依赖性_Count_u教育程度_u个月_不活跃_12 _mon_1\'，

\'天真的_Bayes_分类器_消耗_标志_卡片_类别_联系人_Count_12 _mon_依赖性_Count_受教育程度_u月_不活跃_12_mon_2\']，

dtype=\'object\'
\'\'\'

简单看一下缺失情况

#木有空值，但是不代表没有其他表达类型的空值，比如说\'-\'
BankChurners.isnull().sum()

我们按照数据类型分一下字段

#数据特征
numeric_features = BankChurners.select_dtypes(include=[np.number])
print(numeric_features.columns)
#类别特征
categorical_features = BankChurners.select_dtypes(include=[np.object])
print(categorical_features.columns)

 

Index([\'CLIENTNUM\', \'Customer_Age\', \'Dependent_count\', \'Months_on_book\',
       \'Total_Relationship_Count\', \'Months_Inactive_12_mon\',
       \'Contacts_Count_12_mon\', \'Credit_Limit\', \'Total_Revolving_Bal\',
       \'Avg_Open_To_Buy\', \'Total_Amt_Chng_Q4_Q1\', \'Total_Trans_Amt\',
       \'Total_Trans_Ct\', \'Total_Ct_Chng_Q4_Q1\', \'Avg_Utilization_Ratio\',
       \'Naive_Bayes_Classifier_Attrition_Flag_Card_Category_Contacts_Count_12_mon_Dependent_count_Education_Level_Months_Inactive_12_mon_1\',
       \'Naive_Bayes_Classifier_Attrition_Flag_Card_Category_Contacts_Count_12_mon_Dependent_count_Education_Level_Months_Inactive_12_mon_2\'],
      dtype=\'object\')
Index([\'Attrition_Flag\', \'Gender\', \'Education_Level\', \'Marital_Status\',
       \'Income_Category\', \'Card_Category\'],
      dtype=\'object\')

看看类别型变量的唯一值

for i in list(categorical_features.columns):
    print(\'*************************\')
    print(\'{0}+的唯一值如下：{1}\'.format(i,BankChurners[i].nunique()))
    print(BankChurners[i].unique())

*************************
Attrition_Flag+的唯一值如下：2
[\'Existing Customer\' \'Attrited Customer\']
*************************
Gender+的唯一值如下：2
[\'M\' \'F\']
*************************
Education_Level+的唯一值如下：7
[\'High School\' \'Graduate\' \'Uneducated\' \'Unknown\' \'College\' \'Post-Graduate\'
 \'Doctorate\']
*************************
Marital_Status+的唯一值如下：4
[\'Married\' \'Single\' \'Unknown\' \'Divorced\']
*************************
Income_Category+的唯一值如下：6
[\'$60K - $80K\' \'Less than $40K\' \'$80K - $120K\' \'$40K - $60K\' \'$120K +\'
 \'Unknown\']
*************************
Card_Category+的唯一值如下：4
[\'Blue\' \'Gold\' \'Silver\' \'Platinum\'] 

看来没有乱七八糟的空值

我们在看一下数值型变量的分布

plt.figure(figsize=(20,50))
col = list(numeric_features.columns)
for i in range(1,len(numeric_features.columns)+1):
    ax=plt.subplot(9,2,i) 
    sns.kdeplot(BankChurners[col[i-1]],bw=1.5)
    plt.xlabel(col[i-1])
plt.show()
    

 

 

 我们看上面变量，第一个变量CLIENTNUM 其实就是客户ID，我们可以不用理会

仔细看上面的图片，我们就会发现，其实有些字段其实是类别比较少的数值型变量，但是还是写代码看好，眼睛有时也会出卖自己

for i in list(numeric_features.columns):
    print(\'{0}+的唯一值如下：{1}\'.format(i,BankChurners[i].nunique()))

 

CLIENTNUM+的唯一值如下：10127
Customer_Age+的唯一值如下：45
Dependent_count+的唯一值如下：6
Months_on_book+的唯一值如下：44
Total_Relationship_Count+的唯一值如下：6
Months_Inactive_12_mon+的唯一值如下：7
Contacts_Count_12_mon+的唯一值如下：7
Credit_Limit+的唯一值如下：6205
Total_Revolving_Bal+的唯一值如下：1974
Avg_Open_To_Buy+的唯一值如下：6813
Total_Amt_Chng_Q4_Q1+的唯一值如下：1158
Total_Trans_Amt+的唯一值如下：5033
Total_Trans_Ct+的唯一值如下：126
Total_Ct_Chng_Q4_Q1+的唯一值如下：830
Avg_Utilization_Ratio+的唯一值如下：964
Naive_Bayes_Classifier_Attrition_Flag_Card_Category_Contacts_Count_12_mon_Dependent_count_Education_Level_Months_Inactive_12_mon_1+的唯一值如下：1704
Naive_Bayes_Classifier_Attrition_Flag_Card_Category_Contacts_Count_12_mon_Dependent_count_Education_Level_Months_Inactive_12_mon_2+的唯一值如下：640

我们把它归为一类

num_cate_col = [  \'Dependent_count\', 
       \'Total_Relationship_Count\', \'Months_Inactive_12_mon\',
       \'Contacts_Count_12_mon\', 
       ]

 

剩下的归为一类

numeric_col = [ x for x in list(numeric_features.columns) if x not in num_cate_col]
numeric_col.remove(\'CLIENTNUM\')
numeric_col
\'\'\'
[\'Customer_Age\',
 \'Months_on_book\',
 \'Credit_Limit\',
 \'Total_Revolving_Bal\',
 \'Avg_Open_To_Buy\',
 \'Total_Amt_Chng_Q4_Q1\',
 \'Total_Trans_Amt\',
 \'Total_Trans_Ct\',
 \'Total_Ct_Chng_Q4_Q1\',
 \'Avg_Utilization_Ratio\',
 \'Naive_Bayes_Classifier_Attrition_Flag_Card_Category_Contacts_Count_12_mon_Dependent_count_Education_Level_Months_Inactive_12_mon_1\',
 \'Naive_Bayes_Classifier_Attrition_Flag_Card_Category_Contacts_Count_12_mon_Dependent_count_Education_Level_Months_Inactive_12_mon_2\']
\'\'\'

 

我们在看看类别比较少的数值型变量

for i in num_cate_col:
    print(\'*************************\')
    print(\'{0}+的唯一值如下：{1}\'.format(i,BankChurners[i].nunique()))
    print(BankChurners[i].unique())

 

*************************
Dependent_count+的唯一值如下：6
[3 5 4 2 0 1]
*************************
Total_Relationship_Count+的唯一值如下：6
[5 6 4 3 2 1]
*************************
Months_Inactive_12_mon+的唯一值如下：7
[1 4 2 3 6 0 5]
*************************
Contacts_Count_12_mon+的唯一值如下：7
[3 2 0 1 4 5 6]

看看y值Attrition_Flag

BankChurners[\'Attrition_Flag\'].value_counts().plot.pie(explode=[0,0.1],autopct=\'%1.1f%%\')

 

 

 比例为16%，还好

画类别变量的直方图以及逾期概率图

cate_cold = list(categorical_features.columns)+num_cate_col
for i in cate_cold:
    tmp = pd.crosstab(BankChurners[i], BankChurners[\'Attrition_Flag\'])
    tmp[\'总人数\'] = tmp.sum(axis=1)
    tmp[\'流失率\'] = tmp[\'Attrited Customer\']/tmp[\'总人数\']
    fig, ax1 = plt.subplots()
    ax1.bar(tmp.index,tmp[\'总人数\'],color=\'green\')
    ax2 = ax1.twinx() 
    ax2.plot(tmp.index,tmp[\'流失率\'],color=\'red\')
    plt.show()

 

 

 

 

 

 

 

 构建特征

2021.1.15来补充一下建模的部分

具体过程就不写了，直接附上代码，以及结果

# -*- coding: utf-8 -*-
"""
Created on Fri Jan 15 09:18:46 2021

@author: Administrator
"""

import pandas as pd 
import numpy as np 
import  matplotlib.pyplot as plt
import seaborn as sns
import pycard as pc

#%%导入数据
BankChurners = pd.read_csv(\'D:\\python_home\\预测客户流失\\bankchurners\\BankChurners.csv\')

BankChurners.isnull().sum()

#%%区分数值型变量和类别型变量
num_col = [\'Customer_Age\',
 \'Months_on_book\',
 \'Credit_Limit\',
 \'Total_Revolving_Bal\',
 \'Avg_Open_To_Buy\',
 \'Total_Amt_Chng_Q4_Q1\',
 \'Total_Trans_Amt\',
 \'Total_Trans_Ct\',
 \'Total_Ct_Chng_Q4_Q1\',
 \'Avg_Utilization_Ratio\',
 \'Naive_Bayes_Classifier_Attrition_Flag_Card_Category_Contacts_Count_12_mon_Dependent_count_Education_Level_Months_Inactive_12_mon_1\',
 \'Naive_Bayes_Classifier_Attrition_Flag_Card_Category_Contacts_Count_12_mon_Dependent_count_Education_Level_Months_Inactive_12_mon_2\']


cate_col = [ i for i in list(BankChurners.columns) if i not in num_col]
cate_col.remove(\'CLIENTNUM\')
cate_col
\'\'\'
[\'Attrition_Flag\',
 \'Gender\',
 \'Dependent_count\',
 \'Education_Level\',
 \'Marital_Status\',
 \'Income_Category\',
 \'Card_Category\',
 \'Total_Relationship_Count\',
 \'Months_Inactive_12_mon\',
 \'Contacts_Count_12_mon\']
\'\'\'

#%%cate_col.unqion()
for i in cate_col:
    print(i,BankChurners[i].unique())
\'\'\'
Attrition_Flag [\'Existing Customer\' \'Attrited Customer\']
Gender [\'M\' \'F\']
Dependent_count [3 5 4 2 0 1]
Education_Level [\'High School\' \'Graduate\' \'Uneducated\' \'Unknown\' \'College\' \'Post-Graduate\'
 \'Doctorate\']
Marital_Status [\'Married\' \'Single\' \'Unknown\' \'Divorced\']
Income_Category [\'$60K - $80K\' \'Less than $40K\' \'$80K - $120K\' \'$40K - $60K\' \'$120K +\'
 \'Unknown\']
Card_Category [\'Blue\' \'Gold\' \'Silver\' \'Platinum\']
Total_Relationship_Count [5 6 4 3 2 1]
Months_Inactive_12_mon [1 4 2 3 6 0 5]
Contacts_Count_12_mon [3 2 0 1 4 5 6]
\'\'\'
    
#%%处理目标变量，赋值0和1 流失就是1 
BankChurners.Attrition_Flag = BankChurners.Attrition_Flag.map({\'Existing Customer\':0,\'Attrited Customer\':1})


#%%类别变量的iv计算
cate_iv_woedf = pc.WoeDf()
clf = pc.NumBin()
for i in cate_col[1:]:
    cate_iv_woedf.append(pc.cross_woe(BankChurners[i] ,BankChurners.Attrition_Flag))
cate_iv_woedf.to_excel(\'tmp11\')


#尝试使用二者组合变量，发现并没有作用
BankChurners[\'Gender_Dependent_count\'] = BankChurners.apply(lambda x :x.Gender+ str(x.Dependent_count),axis=1)
pc.cross_woe(BankChurners.Gender_Dependent_count,BankChurners.Attrition_Flag)
BankChurners[\'Education_Level_Income_Category\'] = BankChurners.apply(lambda x :x.Education_Level+ str(x.Income_Category),axis=1)
pc.cross_woe(BankChurners.Education_Level_Income_Category,BankChurners.Attrition_Flag)
BankChurners.pop(\'Gender_Dependent_count\')
BankChurners.pop(\'Education_Level_Income_Category\')

#%%数值变量的iv值计算
num_iv_woedf = pc.WoeDf()
clf = pc.NumBin()
for i in num_col:
    clf.fit(BankChurners[i] ,BankChurners.Attrition_Flag)
    clf.generate_transform_fun()
    num_iv_woedf.append(clf.woe_df_)
num_iv_woedf.to_excel(\'tmp12\')

#%%简单处理一下变量
#Total_Relationship_Count,Months_Inactive_12_mon，Contacts_Count_12_mon,类别型变量就不需要管了，保留原来的分区即可


#Credit_Limit
from numpy import *
BankChurners[\'Credit_Limit_bin\'] = pd.cut(BankChurners.Credit_Limit,bins=[-inf, 1438.65, 1638.5, 1900.5, 2477.5, 3398.5, 5061.5, inf])
BankChurners[\'Total_Revolving_Bal_bin\'] = pd.cut(BankChurners.Total_Revolving_Bal,bins=[-inf, 66.0, 581.5, 979.5, 2512.5, inf])
BankChurners[\'Avg_Open_To_Buy_bin\'] = pd.cut(BankChurners.Avg_Open_To_Buy,bins=[-inf, 447.5, 1038.5, 1437.0, 1944.5, 2229.5,  inf])
BankChurners[\'Total_Amt_Chng_Q4_Q1_bin\'] = pd.cut(BankChurners.Total_Amt_Chng_Q4_Q1,bins=[-inf,  0.3685, 0.4355, 0.5045, 0.5325, 1.0625, inf])
BankChurners[\'Total_Trans_Amt_bin\'] = pd.cut(BankChurners.Total_Trans_Amt,bins=[-inf, 1001.0, 2010.5, 2729.5, 2932.5, 3152.0, 5365.0, 11093.0, inf])
BankChurners[\'Total_Trans_Ct_bin\'] = pd.cut(BankChurners.Total_Trans_Ct,bins=[-inf, 20.5, 37.5, 47.5, 54.5, 57.5, 64.5, 78.5, inf])
BankChurners[\'Total_Ct_Chng_Q4_Q1_bin\'] = pd.cut(BankChurners.Total_Ct_Chng_Q4_Q1,bins=[-inf, 0.4075,  0.4875, 0.504, 0.6015, 0.6565, inf])
BankChurners[\'Avg_Utilization_Ratio_bin\'] = pd.cut(BankChurners.Avg_Utilization_Ratio,bins=[-inf,  0.0205, 0.4505, 0.7985,  inf])
BankChurners[\'Naive_Bayes1_bin\'] = pd.cut(BankChurners.Naive_Bayes_Classifier_Attrition_Flag_Card_Category_Contacts_Count_12_mon_Dependent_count_Education_Level_Months_Inactive_12_mon_1,\
                                         bins=[-inf, 0.4736, inf])
BankChurners[\'Naive_Bayes2_bin\'] = pd.cut(BankChurners.Naive_Bayes_Classifier_Attrition_Flag_Card_Category_Contacts_Count_12_mon_Dependent_count_Education_Level_Months_Inactive_12_mon_2,\
                                         bins=[-inf, 0.5264, inf])


iv_col = [i for i in [\'Total_Relationship_Count\',\'Months_Inactive_12_mon\',\'Contacts_Count_12_mon\',
          ] + list(BankChurners.columns)[-10:]]

cate_iv_woedf = pc.WoeDf()
clf = pc.NumBin()
for i in iv_col:
    cate_iv_woedf.append(pc.cross_woe(BankChurners[i] ,BankChurners.Attrition_Flag))
cate_iv_woedf.to_excel(\'tmp11\')

BankChurners.Contacts_Count_12_mon[BankChurners.Contacts_Count_12_mon ==6] =5

#%%解决方案如下：
#解决方案：后面这两个就不入模型了，作为规则即可，不然会导致过拟合，上面的出现了无穷的变量要合并区间
iv_col = [i for i in [\'Total_Relationship_Count\',\'Months_Inactive_12_mon\',\'Contacts_Count_12_mon\',
          ] + list(BankChurners.columns)[-10:-2]]

cate_iv_woedf = pc.WoeDf()
clf = pc.NumBin()
for i in iv_col:
    cate_iv_woedf.append(pc.cross_woe(BankChurners[i] ,BankChurners.Attrition_Flag))
cate_iv_woedf.to_excel(\'tmp11\')

#%%woe转换
pc.obj_info(cate_iv_woedf)

cate_iv_woedf.bin2woe(BankChurners,iv_col)

model_col = [i for i in [\'CLIENTNUM\', \'Attrition_Flag\']+list(BankChurners.columns)[-11:]]

#%%建模
import pandas as pd
import matplotlib.pyplot as plt #导入图像库
import matplotlib
import seaborn as sns
import statsmodels.api as sm
from sklearn.metrics import roc_curve, auc
from sklearn.model_selection import train_test_split

X = BankChurners[list(BankChurners.columns)[-11:]]
Y = BankChurners[\'Attrition_Flag\']


x_train,x_test,y_train,y_test=train_test_split(X,Y,test_size=0.3,random_state=0)

#(10127, 44)

X1=sm.add_constant(x_train)   #在X前加上一列常数1，方便做带截距项的回归
logit=sm.Logit(y_train.astype(float),X1.astype(float))
result=logit.fit()
result.summary()
result.params



X3 = sm.add_constant(x_test)
resu = result.predict(X3.astype(float))
fpr, tpr, threshold = roc_curve(y_test, resu)
rocauc = auc(fpr, tpr)
plt.plot(fpr, tpr, \'b\', label=\'AUC = %0.2f\' % rocauc)
plt.legend(loc=\'lower right\')
plt.plot([0, 1], [0, 1], \'r--\')
plt.xlim([0, 1])
plt.ylim([0, 1])
plt.ylabel(\'真正率\')
plt.xlabel(\'假正率\')
plt.show()

模型的参数如下：

const -1.662204
Total_Relationship_C_woe -2.307529
Months_Inactive_12_woe -1.050666
Contacts_Count_12_woe -0.742126
Credit_Limit_woe -0.633916
Total_Revolving_Bal_woe -0.914706
Avg_Open_To_Buy_woe -0.120073
Total_Amt_Chng_Q4_Q1_woe -0.845637
Total_Trans_Amt_woe -0.924220
Total_Trans_Ct_woe -0.601850
Total_Ct_Chng_Q4_Q1_woe -0.493193
Avg_Utilization_Ratio_woe -0.265479

测试集的auc图片如下：

0.9651589169400466

 

 看看训练集的auc

resu_1 = result.predict(X1.astype(float))
fpr, tpr, threshold = roc_curve(y_train, resu_1)
rocauc = auc(fpr, tpr)
plt.plot(fpr, tpr, \'b\', label=\'AUC = %0.2f\' % rocauc)
plt.legend(loc=\'lower right\')
plt.plot([0, 1], [0, 1], \'r--\')
plt.xlim([0, 1])
plt.ylim([0, 1])
plt.ylabel(\'真正率\')
plt.xlabel(\'假正率\')
plt.show()

0.9683602097734862

 

 放一下入模变量的iv值

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 总结：

1.变量的逾期率不单调，没有关系，但是分箱时每个箱子不能太小，太小需要合并，最好挑选iv值比较好（2%以上）的入魔，使用woe去训练模型

2.关于iv值为无穷时候的做法：

IVi无论等于负无穷还是正无穷，都是没有意义的。
由上述问题我们可以看到，使用IV其实有一个缺点，就是不能自动处理变量的分组中出现响应比例为0或100%的情况。那么，遇到响应比例为0或者100%的情况，我们应该怎么做呢？建议如下：
（1）如果可能，直接把这个分组做成一个规则，作为模型的前置条件或补充条件；
（2）重新对变量进行离散化或分组，使每个分组的响应比例都不为0且不为100%，尤其是当一个分组个体数很小时（比如小于100个），强烈建议这样做，因为本身把一个分组个体数弄得很小就不是太合理。
（3）如果上面两种方法都无法使用，建议人工把该分组的响应数和非响应的数量进行一定的调整。如果响应数原本为0，可以人工调整响应数为1，如果非响应数原本为0，可以人工调整非响应数为1.

本次也出现了两种iv值为无穷的情况

 

 

 

 

 

 

 我的处理方法是：

解决方案：后面这两个就不入模型了，作为规则即可，不然可能导致过拟合，上面的出现了无穷的变量要合并区间

 还有一些个人信息的特征，比如年龄，婚姻等等，由于iv值很低就没有放进去，但是还是值得探讨一下，不过本次的模型的效果很好，就不做太复杂的模型了

 

但是还有一个疑惑是：

Naive_Bayes_Classifier_Attrition_Flag_Card_Category_Contacts_Count_12_mon_Dependent_count_Education_Level_Months_Inactive_12_mon_1
Naive_Bayes_Classifier_Attrition_Flag_Card_Category_Contacts_Count_12_mon_Dependent_count_Education_Level_Months_Inactive_12_mon_2

最后两个变量实在是太厉害了，不知道是不是别人训练好的y?

需要查资料看看

最后在官网找到了答案