HW9-第七讲：使用集成算法建立个人银行反欺诈模型-

1、背景介绍：

C银行信用卡中心在对欺诈风险和反欺诈技术作了充分研究之后，融合内外部数据，

建立以评分模型为支撑的欺诈识别和防范系统，以满足精细化管理的需要。

本次作业根据“FRAUD_TRAIN_Samp.csv”，引自《SAS金融数据挖掘与建模》,建立信用卡申请反欺诈模型。

2、主要变量说明如下：

#无-组合算法为黑箱模型，无需知道变量含义

3、作业安排：

3.1 基础知识：

      1）比较逻辑回归、决策树、神经网络、组合算法的适用场景和优缺点。

3.2 案例解答步骤如下：

     1）使用决策树、神经网络、组合算法建立反欺诈模型，比较三个模型的表现。

     2)自学绘制PR曲线：横轴为精确度(Precise)，纵轴为召回率(Recall)。

 #%%

 import os

 import pandas as pd

 

 os.chdir(r'D:\Learningfile\天善学院\280_Ben_八大直播八大案例配套课件\提交-第七讲：使用集成算法建立个人银行反欺诈模型\作业')

 data=pd.read_excel('FRAUD_TRAIN_Samp.xlsx')

 #%%

 data.target.value_counts()

 #%%

 data.F81.value_counts()

 #%%

 data.F74.value_counts()

  #%%

 data.F75.value_counts()

  #%%

 data.F80.value_counts()

 #%%

 import numpy as np

 

 dict_F81={'无等级':0,

           np.nan:0,

           '平衡':1,

           '稳健':1,

           '成长':1,

           '进取':1,

           '保守':1}



 data['F81']=data['F81'].map(dict_F81)

 

#%%

dict_F75={'M':1,

        'F':0,

        np.nan:2}

data.F75=data.F75.map(dict_F75)



 #%%缺失值处理

 data1=data.copy()



 data1=data1.drop(['csr_id','F74','F80'],axis=1)



 data1=data1.fillna(0)



 #%%

 from sklearn.cross_validation import train_test_split

 data_X=data1.iloc[:,1:]

 data_y=data1.iloc[:,0]

 train_X,test_X,train_y,test_y=train_test_split(data_X,data_y,test_size=0.4, train_size=0.6, random_state=12345)

 

 #%%



 #采样

 from imblearn.combine import SMOTETomek

 kos = SMOTETomek(random_state=0)  # 综合采样

 X_kos, y_kos = kos.fit_sample(train_X, train_y)



#%% 

#决策树建模

from sklearn.tree import DecisionTreeClassifier

from sklearn import metrics

from sklearn.model_selection import GridSearchCV



clf = DecisionTreeClassifier(criterion='gini', random_state=1234)

param_grid = {'max_depth':[4,12,13,14], 'max_leaf_nodes':[4,12,30,31,32]}

cv = GridSearchCV(clf, param_grid=param_grid, scoring='f1')



cv.fit(X_kos, y_kos)



predict_test = cv.predict(test_X)

test_proba = cv.predict_proba(test_X)[:, 1]

#%%



from sklearn import metrics

fpr_test, tpr_test, th_test = metrics.roc_curve(test_y, test_proba)

print('AUC = %6.4f' %metrics.auc(fpr_test, tpr_test))



#print(metrics.classification_report(test_y, predict_test))

#%%

cv.best_params_



#%%

#神经网络建模

#标准化

from sklearn.preprocessing import MinMaxScaler



scaler = MinMaxScaler()

scaler.fit(train_X)



scaled_train_data = scaler.transform(train_X)

scaled_test_data = scaler.transform(test_X)

# In[5]:

from sklearn.neural_network import MLPClassifier



mlp = MLPClassifier(hidden_layer_sizes=(10,), 

                    activation='logistic', alpha=0.1, max_iter=1000)



mlp.fit(scaled_train_data, train_y)

mlp

#%%

predict_test2 = mlp.predict(scaled_test_data)

test_proba2 = mlp.predict_proba(scaled_test_data)[:, 1]



from sklearn import metrics

fpr_test2, tpr_test2, th_test2 = metrics.roc_curve(test_y, test_proba2)

print('AUC = %6.4f' %metrics.auc(fpr_test2, tpr_test2))

#print(metrics.classification_report(test_y, predict_test2))

#%%

##RF建模

import sklearn.ensemble as ensemble

from sklearn.ensemble import RandomForestClassifier

param_grid = {

    'criterion':['entropy','gini'],

    'max_depth':[4,5,6,7],

    'n_estimators':[24,26,28],  #决策树个数-随机森林特有参数

    'max_features':[0.02,0.04,0.4], #每棵决策树使用的变量占比-随机森林特有参数

    'min_samples_split':[74,75,78] 

}



rfc = RandomForestClassifier()

rfccv = GridSearchCV(estimator=rfc, param_grid=param_grid, scoring='roc_auc', cv=4,n_jobs=-1)

rfccv.fit(train_X, train_y)

test_est3 = rfccv.predict(test_X)

test_proba3 = rfccv.predict_proba(test_X)[:, 1]

fpr_test3,tpr_test3,th_test3=metrics.roc_curve(test_y,test_proba3)

print('RF ROC')

print('ROC=%.4f' %metrics.auc(fpr_test3,tpr_test3))

#%%

rfccv.best_params_







#%%

#首先对n_estimators进行网格搜索  

param_test1= {'n_estimators':range(10,71,10)}  

gsearch1= GridSearchCV(estimator =RandomForestClassifier(min_samples_split=100,  

                                 min_samples_leaf=20,max_depth=8,max_features='sqrt' ,random_state=10),  

                       param_grid =param_test1, scoring='roc_auc',cv=5)  

gsearch1.fit(train_X, train_y)  

gsearch1.grid_scores_,gsearch1.best_params_, gsearch1.best_score_ 



#%%

#这样我们得到了最佳的弱学习器迭代次数，接着我们对决策树最大深度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(2,8,1), 'min_samples_split':range(3,21,5)}  

gsearch2= GridSearchCV(estimator = ensemble.RandomForestClassifier(n_estimators= 40,  

                                 min_samples_leaf=20,max_features='sqrt' ,oob_score=True,random_state=10),  

   param_grid = param_test2,scoring='roc_auc',iid=False, cv=5)  

gsearch2.fit(train_X, train_y)  

gsearch2.grid_scores_,gsearch2.best_params_, gsearch2.best_score_  

#%%



rfc.fit(train_X,train_y)

print(rfc.oob_score_)

#%%

#已经取了三个最优参数，看看现在模型的袋外分数：  



rf1= RandomForestClassifier(n_estimators= 40, max_depth=3, min_samples_split=3,  

                                 min_samples_leaf=20,max_features='sqrt' ,oob_score=True,random_state=10)  

rf1.fit(train_X, train_y)  

print(rf1.oob_score_)

#输出结果为：0.9485  

#可见此时我们的袋外分数有一定的提高。也就是时候模型的泛化能力增强了。

#对于内部节点再划分所需最小样本数min_samples_split，我们暂时不能一起定下来，因为这个还和决策树其他的参数存在关联。

#下面我们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。  

#%%   

#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(3,30,5), 'min_samples_leaf':range(10,81,10)}  

gsearch3= GridSearchCV(estimator = RandomForestClassifier(n_estimators= 40,max_depth=3,  

                                 max_features='sqrt' ,oob_score=True, random_state=10),  

   param_grid = param_test3,scoring='roc_auc',iid=False, cv=5)  

gsearch3.fit(train_X, train_y)  

gsearch3.grid_scores_,gsearch2.best_params_, gsearch2.best_score_  



#%%

#最后我们再对最大特征数max_features做调参:  

param_test4= {'max_features':range(3,11,2)}  

gsearch4= GridSearchCV(estimator = RandomForestClassifier(n_estimators= 40,max_depth=3, min_samples_split=3,  

                                 min_samples_leaf=50 ,oob_score=True, random_state=10),  

   param_grid = param_test4,scoring='roc_auc',iid=False, cv=5)  

gsearch4.fit(train_X, train_y)  

gsearch4.grid_scores_,gsearch4.best_params_, gsearch4.best_score_  



#%%

#用我们搜索到的最佳参数，我们再看看最终的模型拟合：  

rf2= RandomForestClassifier(n_estimators= 40, max_depth=3, min_samples_split=3,  

                                 min_samples_leaf=50,max_features=5 ,oob_score=True, random_state=10)  

rf2.fit(train_X, train_y)  

print(rf2.oob_score_)

#此时的输出为：0.9485  

#可见此时模型的袋外分数基本没有提高，主要原因是0.9485已经是一个很高的袋外分数了，如果想进一步需要提高模型的泛化能力，我们需要更多的数据。  



#%%

test_est31 = rf2.predict(test_X)

test_proba31 = rf2.predict_proba(test_X)[:, 1]

fpr_test31,tpr_test31,th_test31=metrics.roc_curve(test_y,test_proba31)

print('rf ROC')

print('ROC=%.4f' %metrics.auc(fpr_test31,tpr_test31))



#%%

#Adaboost算法

param_grid = {

    #'base_estimator':['DecisionTreeClassifier'],

    'learning_rate':[0.1,0.3,0.5,0.7,1]

}



abc = ensemble.AdaBoostClassifier(n_estimators=100,algorithm='SAMME')

abccv = GridSearchCV(estimator=abc, param_grid=param_grid, scoring='roc_auc', cv=4,n_jobs=-1)

abccv.fit(train_X, train_y)

test_est4 = abccv.predict(test_X)

test_proba4 = abccv.predict_proba(test_X)[:, 1]

print("abc classifier accuracy:")

print(metrics.classification_report(test_y,test_est4))



fpr_test4,tpr_test4,th_test4=metrics.roc_curve(test_y,test_proba4)

print('adaboost ROC')

print('ROC=%.4f' %metrics.auc(fpr_test4,tpr_test4))

#%%

abccv.best_params_



# In[ ]:

#GBDT

param_grid = {

    'loss':['deviance','exponential'],

    'learning_rate':[0.1,0.3,0.8],

    'n_estimators':[26,28,30,32],  #决策树个数-GBDT特有参数

    'max_depth':[4,5,6],  #单棵树最大深度-GBDT特有参数

    'min_samples_split':[2,15,20,25] 

}



gbc = ensemble.GradientBoostingClassifier()

gbccv = GridSearchCV(estimator=gbc, param_grid=param_grid, scoring='roc_auc', cv=4,n_jobs=-1)

gbccv.fit(train_X, train_y)

test_est5 = gbccv.predict(test_X)

test_proba5 = gbccv.predict_proba(test_X)[:, 1]

print("gradient boosting accuracy:")

print(metrics.classification_report(test_y,test_est5))



fpr_test5,tpr_test5,th_test5=metrics.roc_curve(test_y,test_proba5)

print('gbdt ROC')

print('ROC=%.4f' %metrics.auc(fpr_test5,tpr_test5))

#%%

gbccv.best_params_





#%%

#PR曲线

from sklearn.metrics import precision_recall_curve

import matplotlib.pyplot as plt



precision,recall,threshold=precision_recall_curve(test_y, test_proba3)



plt.figure(figsize=[6,6])

plt.plot(precision,recall)

plt.ylabel('recall')

plt.xlabel('precision')

plt.title('pr curve')

plt.show()