原理很简单，初始分20箱或更多，先确保每箱中都含有0，1标签，对不包含0，1标签的箱向前合并，计算各箱卡方值，对卡方值最小的箱向后合并，代码如下

import pandas as pd

import numpy as np

import scipy

from scipy import stats

def chi_bin(DF,var,target,binnum=5,maxcut=20):

'''

DF:data

var:variable

target:target / label

binnum: the number of bins output

maxcut: initial bins number

'''

data=DF[[var,target]]

#equifrequent cut the var into maxcut bins

data["cut"],breaks=pd.qcut(data[var],q=maxcut,duplicates="drop",retbins=True)

#count 1,0 in each bin

count_1=data.loc[data[target]==1].groupby("cut")[target].count()

count_0=data.loc[data[target]==0].groupby("cut")[target].count()

#get bins value: min,max,count 0,count 1

bins_value=[*zip(breaks[:maxcut-1],breaks[1:],count_0,count_1)]

#define woe

def woe_value(bins_value):

df_woe=pd.DataFrame(bins_value)

df_woe.columns=["min","max","count_0","count_1"]

df_woe["total"]=df_woe.count_1+df_woe.count_0

df_woe["bad_rate"]=df_woe.count_1/df_woe.total

df_woe["woe"]=np.log((df_woe.count_0/df_woe.count_0.sum())/(df_woe.count_1/df_woe.count_1.sum()))

return df_woe

#define iv

def iv_value(df_woe):

rate=(df_woe.count_0/df_woe.count_0.sum())-(df_woe.count_1/df_woe.count_1.sum())

iv=np.sum(rate * df_woe.woe)

return iv

#make sure every bin contain 1 and 0

##first bin merge backwards

for i in range(len(bins_value)):

if 0 in bins_value[0][2:]:

bins_value[0:2]=[(

bins_value[0][0],

bins_value[1][1],

bins_value[0][2]+bins_value[1][2],

bins_value[0][3]+bins_value[1][3])]

continue

##bins merge forwards

if 0 in bins_value[i][2:]:

bins_value[i-1:i+1]=[(

bins_value[i-1][0],

bins_value[i][1],

bins_value[i-1][2]+bins_value[i][2],

bins_value[i-1][3]+bins_value[i][3])]

break

else:

break

#calculate chi-square merge the minimum chisquare

while len(bins_value)>binnum:

chi_squares=[]

for i in range(len(bins_value)-1):

a=bins_value[i][2:]

b=bins_value[i+1][2:]

chi_square=scipy.stats.chi2_contingency([a,b])[0]

chi_squares.append(chi_square)

#merge the minimum chisquare backwards

i = chi_squares.index(min(chi_squares))

bins_value[i:i+2]=[(

bins_value[i][0],

bins_value[i+1][1],

bins_value[i][2]+bins_value[i+1][2],

bins_value[i][3]+bins_value[i+1][3])]

df_woe=woe_value(bins_value)

#print bin number and iv

print("箱数：{},iv:{:.6f}".format(len(bins_value),iv_value(df_woe)))

#return bins and woe information

return woe_value(bins_value)

以下是效果：

初始分成10箱，目标为3箱

chi_bin(data,"age","SeriousDlqin2yrs",binnum=3,maxcut=10)

箱数：8,iv:0.184862

箱数：7,iv:0.184128

箱数：6,iv:0.179518

箱数：5,iv:0.176980

箱数：4,iv:0.172406

箱数：3,iv:0.160015

min max count_0 count_1 total bad_rate woe

0 0.0 52.0 70293 7077 77370 0.091470 -0.266233

1 52.0 61.0 29318 1774 31092 0.057056 0.242909

2 61.0 72.0 26332 865 27197 0.031805 0.853755