决策树、SVM、随机森林在评定信用等级上的应用

以下为我们这次的数据集信息，分别是各类特征和信用评定Label，属于二分类问题。

本文章想通过比较决策树、SVM和随机森林在该数据集上的表现

导入数据，查看缺失值

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

data =pd.read_excel('./GermanCredit.xls', sheet_name='Data')  #读取xls文件的Data sheet
data.head()

num_features = ['DURATION','AMOUNT','INSTALL_RATE','AGE','NUM_CREDITS','NUM_DEPENDENTS']
cat_features = data.columns.drop(num_features + ['OBS#'])

data.isnull().sum()
# 都没有缺失值

OBS#                0
CHK_ACCT            0
DURATION            0
HISTORY             0
NEW_CAR             0
USED_CAR            0
FURNITURE           0
RADIO/TV            0
EDUCATION           0
RETRAINING          0
AMOUNT              0
SAV_ACCT            0
EMPLOYMENT          0
INSTALL_RATE        0
MALE_DIV            0
MALE_SINGLE         0
MALE_MAR_or_WID     0
CO-APPLICANT        0
GUARANTOR           0
PRESENT_RESIDENT    0
REAL_ESTATE         0
PROP_UNKN_NONE      0
AGE                 0
OTHER_INSTALL       0
RENT                0
OWN_RES             0
NUM_CREDITS         0
JOB                 0
NUM_DEPENDENTS      0
TELEPHONE           0
FOREIGN             0
RESPONSE            0
dtype: int64

将连续特征离散化

发现DURATION是贷款期限，分布在4-72个月之间，而且分布是一个看似左偏的正态分布，做一个hist图看得更清楚！

plt.hist(data['DURATION'])

(array([171., 262., 337.,  57.,  86.,  17.,  54.,   2.,  13.,   1.]),
 array([ 4. , 10.8, 17.6, 24.4, 31.2, 38. , 44.8, 51.6, 58.4, 65.2, 72. ]),
 <a list of 10 Patch objects>)

取五分位数，将DURATION特征转化成cat_features做离散化处理

x<20 dua_rank = 1
20<x<40 dua_rank = 2
40<x<60 dua_rank = 3
60<x<72 dua_rank = 4
并且创造一个新特征 dua_rank 添加在new_data中,也可以用sklearn.KBinsDiscretizer进行分箱处理

dua_rank = []
duration = data['DURATION']
for i in duration:
    if i <=20:
        dua_rank.append(1)
    elif i<= 40:
        dua_rank.append(2)
    elif i < 60:
        dua_rank.append(3)
    else:
        dua_rank.append(4)

可以看出，大部分的duration分布在rank1、2的区间内

plt.hist(dua_rank,bins = 4)

(array([554., 365.,  67.,  14.]),
 array([1.  , 1.75, 2.5 , 3.25, 4.  ]),
 <a list of 4 Patch objects>)

new_data = data.copy()
new_data['dua_rank'] = dua_rank

new_data.head()

	OBS#	CHK_ACCT	DURATION	HISTORY	NEW_CAR	USED_CAR	FURNITURE	RADIO/TV	EDUCATION	RETRAINING	...	OTHER_INSTALL	RENT	OWN_RES	NUM_CREDITS	JOB	NUM_DEPENDENTS	TELEPHONE	FOREIGN	RESPONSE	dua_rank
0	1	0	6	4	0	0	0	1	0	0	...	0	0	1	2	2	1	1	0	1	1
1	2	1	48	2	0	0	0	1	0	0	...	0	0	1	1	2	1	0	0	0	3
2	3	3	12	4	0	0	0	0	1	0	...	0	0	1	1	1	2	0	0	1	1
3	4	0	42	2	0	0	1	0	0	0	...	0	0	0	1	2	2	0	0	1	3
4	5	0	24	3	1	0	0	0	0	0	...	0	0	0	2	2	2	0	0	0	2

5 rows × 33 columns

plt.hist(data['AMOUNT'])

(array([445., 293.,  97.,  80.,  38.,  19.,  14.,   8.,   5.,   1.]),
 array([  250. ,  2067.4,  3884.8,  5702.2,  7519.6,  9337. , 11154.4,
        12971.8, 14789.2, 16606.6, 18424. ]),
 <a list of 10 Patch objects>)

我们也将AMOUNT特征分为1-10级，用十分位点作为评分标准

同样可以用sklearn.KBinsDiscretizer进行分箱离散化

percent = np.percentile(data['AMOUNT'], [i * 10 for i in range(1,10)])

amount_rank = []
for i in data['AMOUNT']:
    if i < percent[0]:
        amount_rank