（十三）GBDT模型用于评分卡模型python实现

import pandas as pd
import time
import numpy as np
import re
from sklearn.ensemble import GradientBoostingClassifier
from sklearn import cross_validation,metrics
from sklearn.model_selection import GridSearchCV,train_test_split
import matplotlib.pylab as plt
import datetime
from dateutil.relativedelta import relativedelta
from numpy import log
from sklearn.metrics import roc_auc_score
from sklearn.feature_extraction import DictVectorizer
from sklearn.preprocessing import OneHotEncoder
from sklearn.linear_model.logistic import LogisticRegression
 

def CareerYear(x):
    #对工作年限进行转换
    if str(x).find(‘nan‘) > -1:
        return -1
    elif str(x).find("10+")>-1:   #将"10＋years"转换成 11
        return 11
    elif str(x).find(‘< 1‘) > -1:  #将"< 1 year"转换成 0
        return 0
    else:
        return int(re.sub("D","",x))   #其余数据，去掉"years"并转换成整数
 
 
def DescExisting(x):
    #将desc变量转换成有记录和无记录两种
    if type(x).__name__ == ‘float‘:
        return ‘no desc‘
    else:
        return ‘desc‘
 
 
def ConvertDateStr(x):
    mth_dict = {‘Jan‘: 1,‘Feb‘: 2,‘Mar‘: 3,‘Apr‘: 4,‘May‘: 5,‘Jun‘: 6,‘Jul‘: 7,‘Aug‘: 8,‘Sep‘: 9,‘Oct‘: 10,‘Nov‘: 11,‘Dec‘: 12}
    if str(x) == ‘nan‘:
        return datetime.datetime.fromtimestamp(time.mktime(time.strptime(‘9900-1‘,‘%Y-%m‘)))
        #time.mktime 不能读取1970年之前的日期
    else:
        yr = int(x[4:6])
        if yr <=17:
            yr = 2000+yr
        else:
            yr = 1900 + yr
        mth = mth_dict[x[:3]]
        return datetime.datetime(yr,mth,1)
 
 
def MonthGap(earlyDate,lateDate):
    if lateDate > earlyDate:
        gap = relativedelta(lateDate,earlyDate)
        yr = gap.years
        mth = gap.months
        return yr*12+mth
    else:
        return 0
 
 
def MakeupMissing(x):
    if np.isnan(x):
        return -1
    else:
        return x
 
 
 
‘‘‘
第一步：数据准备
‘‘‘
folderOfData = foldOfData = ‘H:/‘
allData = pd.read_csv(folderOfData + ‘数据集.csv‘,header = 0,encoding = ‘latin1‘,engine =‘python‘)
allData[‘term‘] = allData[‘term‘].apply(lambda x: int(x.replace(‘ months‘,‘‘)))
# 处理标签：Fully Paid是正常用户；Charged Off是违约用户
allData[‘y‘] = allData[‘loan_status‘].map(lambda x: int(x == ‘Charged Off‘))
 
‘‘‘
由于存在不同的贷款期限（term），申请评分卡模型评估的违约概率必须要在统一的期限中，且不宜太长，所以选取term＝36months的行本
‘‘‘
allData1 = allData.loc[allData.term == 36]
trainData,testData = train_test_split(allData1,test_size=0.4)
 
 
 
‘‘‘
第二步：数据预处理
‘‘‘
# 将带％的百分比变为浮点数
trainData[‘int_rate_clean‘] = trainData[‘int_rate‘].map(lambda x: float(x.replace(‘%‘,‘‘))/100)
# 将工作年限进行转化，否则影响排序
trainData[‘emp_length_clean‘] = trainData[‘emp_length‘].map(CareerYear)
# 将desc的缺失作为一种状态，非缺失作为另一种状态
trainData[‘desc_clean‘] = trainData[‘desc‘].map(DescExisting)
# 处理日期。earliest_cr_line的格式不统一，需要统一格式且转换成python的日期
trainData[‘app_date_clean‘] = trainData[‘issue_d‘].map(lambda x: ConvertDateStr(x))
trainData[‘earliest_cr_line_clean‘] = trainData[‘earliest_cr_line‘].map(lambda x: ConvertDateStr(x))
# 处理mths_since_last_delinq。注意原始值中有0，所以用－1代替缺失
trainData[‘mths_since_last_delinq_clean‘] = trainData[‘mths_since_last_delinq‘].map(lambda x:MakeupMissing(x))
trainData[‘mths_since_last_record_clean‘] = trainData[‘mths_since_last_record‘].map(lambda x:MakeupMissing(x))
trainData[‘pub_rec_bankruptcies_clean‘] = trainData[‘pub_rec_bankruptcies‘].map(lambda x:MakeupMissing(x))
 
‘‘‘
第三步：变量衍生
‘‘‘
# 考虑申请额度与收入的占比
trainData[‘limit_income‘] = trainData.apply(lambda x: x.loan_amnt / x.annual_inc,axis = 1)
# 考虑earliest_cr_line到申请日期的跨度，以月份记
trainData[‘earliest_cr_to_app‘] = trainData.apply(lambda x: MonthGap(x.earliest_cr_line_clean,x.app_date_clean),axis = 1)
 
 
‘‘‘
对于类别型变量，需要onehot（独热）编码，再训练GBDT模型
‘‘‘
num_features = [‘int_rate_clean‘,‘emp_length_clean‘,‘annual_inc‘,‘dti‘,‘delinq_2yrs‘,‘earliest_cr_to_app‘,‘inq_last_6mths‘,‘mths_since_last_record_clean‘,‘mths_since_last_delinq_clean‘,‘open_acc‘,‘pub_rec‘,‘total_acc‘,‘limit_income‘,‘earliest_cr_to_app‘]
cat_features = [‘home_ownership‘,‘verification_status‘,‘desc_clean‘,‘purpose‘,‘zip_code‘,‘addr_state‘,‘pub_rec_bankruptcies_clean‘]
 
#独热编码
v = DictVectorizer(sparse=False)
X1 = v.fit_transform(trainData[cat_features].to_dict(‘records‘))
#将独热编码和数值型变量放在一起进行模型训练
X2 = np.matrix(trainData[num_features])
X = np.hstack([X1,X2])
y = trainData[‘y‘]
 
# 未经调参进行GBDT模型训练
gbm0 = GradientBoostingClassifier(random_state=10)
gbm0.fit(X,y)
 
 
 
y_pred = gbm0.predict(X)
y_predprob = gbm0.predict_proba(X)[:,1].T
print("Accuracy : %.4g" % metrics.accuracy_score(y,y_pred))
print("AUC Score (Train): %f" % metrics.roc_auc_score(np.array(y.T),y_predprob))
 
 
 
‘‘‘
第四步：在测试集上测试模型的性能
‘‘‘
# 将带％的百分比变为浮点数
testData[‘int_rate_clean‘] = testData[‘int_rate‘].map(lambda x: float(x.replace(‘%‘,‘‘))/100)
# 将工作年限进行转化，否则影响排序
testData[‘emp_length_clean‘] = testData[‘emp_length‘].map(CareerYear)
# 将desc的缺失作为一种状态，非缺失作为另一种状态
testData[‘desc_clean‘] = testData[‘desc‘].map(DescExisting)
# 处理日期。earliest_cr_line的格式不统一，需要统一格式且转换成python的日期
testData[‘app_date_clean‘] = testData[‘issue_d‘].map(lambda x: ConvertDateStr(x))
testData[‘earliest_cr_line_clean‘] = testData[‘earliest_cr_line‘].map(lambda x: ConvertDateStr(x))
# 处理mths_since_last_delinq。注意原始值中有0，所以用－1代替缺失
testData[‘mths_since_last_delinq_clean‘] = testData[‘mths_since_last_delinq‘].map(lambda x:MakeupMissing(x))
testData[‘mths_since_last_record_clean‘] = testData[‘mths_since_last_record‘].map(lambda x:MakeupMissing(x))
testData[‘pub_rec_bankruptcies_clean‘] = testData[‘pub_rec_bankruptcies‘].map(lambda x:MakeupMissing(x))
 
 
# 考虑申请额度与收入的占比
testData[‘limit_income‘] = testData.apply(lambda x: x.loan_amnt / x.annual_inc,axis = 1)
# 考虑earliest_cr_line到申请日期的跨度，以月份记
testData[‘earliest_cr_to_app‘] = testData.apply(lambda x: MonthGap(x.earliest_cr_line_clean,axis = 1)
 
#用训练集里的onehot编码方式进行编码
X1_test = v.transform(testData[cat_features].to_dict(‘records‘))
X2_test = np.matrix(testData[num_features])
X_test = np.hstack([X1_test,X2_test])
y_test = np.matrix(testData[‘y‘]).T
 
 
### 计算KS值
def KS(df,score,target):
    ‘‘‘
    :param df: 包含目标变量与预测值的数据集,dataframe
    :param score: 得分或者概率,str
    :param target: 目标变量,str
    :return: KS值
    ‘‘‘
    total = df.groupby([score])[target].count()
    bad = df.groupby([score])[target].sum()
    all = pd.DataFrame({‘total‘:total,‘bad‘:bad})
    all[‘good‘] = all[‘total‘] - all[‘bad‘]
    all[score] = all.index
    all = all.sort_values(by=score,ascending=False)
    all.index = range(len(all))
    all[‘badCumRate‘] = all[‘bad‘].cumsum() / all[‘bad‘].sum()
    all[‘goodCumRate‘] = all[‘good‘].cumsum() / all[‘good‘].sum()
    KS = all.apply(lambda x: x.badCumRate - x.goodCumRate,axis=1)
    return max(KS)
 
 
#在测试集上测试GBDT性能
y_pred = gbm0.predict(X_test)
y_predprob = gbm0.predict_proba(X_test)[:,1].T
testData[‘predprob‘] = list(y_predprob)
print("Accuracy : %.4g" % metrics.accuracy_score(y_test,y_pred))
print("AUC Score (Test): %f" % metrics.roc_auc_score(np.array(y_test)[:,0],y_predprob))
print("KS is :%f" % KS(testData,‘predprob‘,‘y‘))
 
 
‘‘‘
GBDT调参
‘‘‘
# 1,选择较小的步长(learning rate)后，对迭代次数(n_estimators)进行调参
 
X = pd.DataFrame(X)
 
param_test1 = {‘n_estimators‘:range(80,81,10)}
gsearch1 = GridSearchCV(estimator = GradientBoostingClassifier(learning_rate=0.1,min_samples_split=30,min_samples_leaf=5,max_depth=8,max_features=‘sqrt‘,subsample=0.8,random_state=10),param_grid = param_test1,scoring=‘roc_auc‘,iid=False,cv=5)
gsearch1.fit(X,y)
gsearch1.best_params_,gsearch1.best_score_
best_n_estimator = gsearch1.best_params_[‘n_estimators‘]
 
 
# 2,对决策树最大深度max_depth和内部节点再划分所需最小样本数min_samples_split进行网格搜索
param_test2 = {‘max_depth‘:range(3,4),‘min_samples_split‘:range(6,7)}
gsearch2 = GridSearchCV(estimator = GradientBoostingClassifier(learning_rate=0.1,n_estimators=best_n_estimator,min_samples_leaf=20,param_grid = param_test2,cv=5)
gsearch2.fit(X,y)
gsearch2.best_params_,gsearch2.best_score_
best_max_depth = gsearch2.best_params_[‘max_depth‘]
 
#3,再对内部节点再划分所需最小样本数min_samples_split和叶子节点最少样本数min_samples_leaf一起调参
param_test3 = {‘min_samples_split‘:range(80,10),‘min_samples_leaf‘:range(50,51,5)}
gsearch3 = GridSearchCV(estimator = GradientBoostingClassifier(learning_rate=0.1,max_depth=best_max_depth,param_grid = param_test3,cv=5)
gsearch3.fit(X,y)
gsearch3.best_params_,gsearch3.best_score_
best_min_samples_split,best_min_samples_leaf = gsearch3.best_params_[‘min_samples_split‘],gsearch3.best_params_[‘min_samples_leaf‘]
 
#4,对最大特征数max_features进行网格搜索
param_test4 = {‘max_features‘:range(int(np.sqrt(X.shape[0])),int(np.sqrt(X.shape[0]))+1,5)}
gsearch4 = GridSearchCV(estimator = GradientBoostingClassifier(learning_rate=0.1,min_samples_leaf =best_min_samples_leaf,min_samples_split =best_min_samples_split,param_grid = param_test4,cv=5)
gsearch4.fit(X,y)
gsearch4.best_params_,gsearch4.best_score_
best_max_features = gsearch4.best_params_[‘max_features‘]
 
#5,对采样比例进行网格搜索
param_test5 = {‘subsample‘:[0.6+i*0.05 for i in range(1)]}
gsearch5 = GridSearchCV(estimator = GradientBoostingClassifier(learning_rate=0.1,max_features=best_max_features,param_grid = param_test5,cv=5)
gsearch5.fit(X,y)
gsearch5.best_params_,gsearch5.best_score_
best_subsample = gsearch5.best_params_[‘subsample‘]
 
 
gbm_best = GradientBoostingClassifier(learning_rate=0.1,subsample =best_subsample,random_state=10)
gbm_best.fit(X,y)
 
 
#在测试集上测试并计算性能
y_pred = gbm_best.predict(X_test)
y_predprob = gbm_best.predict_proba(X_test)[:,1].T
testData[‘predprob‘] = list(y_predprob)
#准确性
print("Accuracy : %.4g" % metrics.accuracy_score(y_test,y_predprob))
print("KS is :%f"%KS(testData,‘y‘))
 
 
###########概率转换为分数########################
def Prob2Score(prob,basePoint,PDO):
    #将概率转化成分数且为正整数
    y = np.log(prob/(1-prob))
    return int(basePoint+PDO/np.log(2)*(-y))
 
basePoint = 250
PDO = 50
testData[‘score‘] = testData[‘predprob‘].map(lambda x:Prob2Score(x,PDO))
testData = testData.sort_values(by = ‘score‘)
#画出分布图
plt.hist(testData[‘score‘],100)
plt.xlabel(‘score‘)
plt.ylabel(‘freq‘)
plt.title(‘distribution‘)