risk-model/example.py at master · huangxianyang/risk-model · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
# -*- coding: utf-8 -*-
"""
评分卡示例:
0.数据探索分析,1.数据预处理,2.特征分箱,3.特征选择,4.模型训练,5.评分卡构建,6.模型评估,7.风险决策
"""

import scorecardpy as sc
import warnings
warnings.filterwarnings('ignore')
import joblib
import riskModel as rs
from  sklearn.model_selection import train_test_split

path = './result/'
# 导入数据
germanCredit = rs.Germancredit()
df = germanCredit.get_data()
germanCredit.get_describe()
df.sample(5)
# 数值化
for col in germanCredit.sub_col:
    df[col] = df[col].apply(lambda x:int(str(x)[1:]))
# 替换
df = df.replace(germanCredit.rep_dict)
df['target'] = df['target'].replace({1:0,2:1})

all_cols = germanCredit.all_feature
num_cols = germanCredit.num_col
cat_cols = germanCredit.cat_col

#######################################
# 0.数据探索分析
#######################################
bs = rs.BasicStat(df=df)
html_report = bs.get_report()
# 保存
with open(path+'dt_report.html','w',encoding='utf8') as f:
    f.write(html_report)

df_report = bs.get_describe()
df_report.to_excel(path+'df_report.xlsx')

#######################################
# 1.数据清洗
#######################################
pr = rs.Preprocess()
str_s = pr.find_str(df=df,cols=num_cols)
print('字符:',str_s)
sign_s = pr.find_sign(df=df,cols=all_cols)
print('特殊符号:',sign_s)
# 离群值
outliers = pr.outlier_drop(df=df,cols=all_cols,low_percent=0.001,up_percent=0.999,cat_percent=0.01)
print('离群值:',outliers)
# 离群值处理
for k,v in outliers.items():
    if type(v) == dict:
        df[k] = df[k].apply(lambda x:v.get('low') if x < v.get('low') else v.get('up') if x > v.get('up') else x)
    elif type(v) == list:
        df[k] = df[k].replace(v,df[k].mode()[0])

# 缺失统计
missm_result = pr.miss_mode_cnt(df=df,cols=all_cols)
print(missm_result)
# 连续变量删除缺失80%+,众数率90%+
# 分类变量删除缺失率90%,众数率95%+
num_cols = [i for i in num_cols if missm_result['miss'][i] < 0.8 and missm_result['mode'][i] < 0.8]
cat_cols = [i for i in cat_cols if missm_result['miss'][i] < 0.8 and missm_result['mode'][i] < 0.95]
print(f'drop num cols :{[i for i in all_cols if i not in num_cols+cat_cols]}')
all_cols = num_cols + cat_cols
# 缺失填充
df = pr.fill_nan(df=df,cols=all_cols,values=-99999,min_miss_rate=0.05)

# 字符变量数值化
df,factorize_dict = pr.factor_map(df=df,cols=cat_cols)
# 过采样
int_cols = cat_cols + [i for i in num_cols if i in germanCredit.int_col]
print('int_cols:',int_cols)

ds = rs.DataSample(df=df,cols=all_cols,target='target',sampling_strategy='minority')
df_res = ds.over_sample(int_cols=int_cols)
# 分类变量还原
to_factorize = {k: {i:j for j,i in v.items()} for k,v in factorize_dict.items()}
df_res = df_res.replace(to_factorize)
df = df.replace(to_factorize)
########################################
# 2.特征工程
########################################
# 样本划分
# woe bin
train,valid = train_test_split(df_res,test_size=0.3,random_state=0)
cat_bin,cat_iv = rs.cat_bin(df=train,cols=cat_cols,target='target',specials=['-99999'],
                            bin_num_limit=5,count_distr_limit=0.05,method='chimerge')
num_bin,num_iv = rs.num_bin(df=train,cols=num_cols,target='target',specials=[-99999],
                            bin_num_limit=8,count_distr_limit=0.05,sc_method='chimerge',
                            non_mono_cols=['age_in_years'],init_bins=15,init_min_samples=0.05,init_method='chi')

bins = {**cat_bin,**num_bin}
ivs = {**cat_iv,**num_iv}
# woe转换
train_woe = sc.woebin_ply(dt=train,bins=bins)
valid_woe = sc.woebin_ply(dt=valid,bins=bins)
df_woe = sc.woebin_ply(dt=df,bins=bins)

#########################################
# 3.特征选择
#########################################
sf = rs.SelectFeature()
high_iv = sf.baseOn_iv(ivd=ivs,thred=0.05,is_draw=False)
low_vif = sf.baseOn_collinear(df=train_woe,high_iv=high_iv,thred=0.7,is_draw=False)
ml_cols, best_C = sf.baseOn_l1(X=train_woe[low_vif.keys()],y=train_woe['target'],Kfold=5,drop_plus=False)

#########################################
# 4.模型训练
#########################################

lr = rs.Lr(C=best_C)
lr.fiting(X=train_woe[ml_cols.keys()],y=train_woe['target'],filename=path)
joblib.dump(lr,path+'lr.pkl')

########################################
# 5.评分卡构建
########################################

card = rs.ScoreCard(lr=lr,bins=bins,ml_cols=[i.replace('_woe','') for i in ml_cols.keys()],
                    score0=600,pdo=50)
card_df = card.score_card()
joblib.dump(card,path+'card.pkl')
########################################
# 6.模型评估,真实样本评估
#######################################
valid['prob'] = lr.predict_proba(valid_woe[ml_cols.keys()])[:,1]
df['prob'] = lr.predict_proba(df_woe[ml_cols.keys()])[:,1]

print(f'valid evaluate:{rs.model_norm(valid["target"],valid["prob"])}')
print(f'test evaluate:{rs.model_norm(df["target"],df["prob"])}')

valid['score'] = card.apply_score(valid)
df['score'] = card.apply_score(df)

########################################
# 7.风险决策
#######################################
cut_off = rs.stragety_score(df=df,step=50,score='score',label='target')
cut_off.to_excel(path+'cut_off.xlsx')