银行客户流失预测

针对银行客户流失预测，主要流程分为：特征预处理、特征选择，分类模型选择与训练。主要工作如下：

1：特征预处理与选择

对性别进行哑变量处理；

对是否有****信息将布尔值转换01表示；

画出年龄直方图可以看出大致呈正态分布，对年龄分段处理后缺失值采用插补方式；

资产当前总额=存储类资产当前总额=本币存储当前总金额   月日均余额=存储类资产月日均余额=本币存储月日均余额  分别删除其中两项；

针对*NUM,*DUR,*AMT,*BAL字段分别进行特征提取（SelectKBest）达到降维效果；

最后整合数据，特征标准化处理最终为44个特征（StandardScaler）。

2：分类模型选择与训练

数据集划分：采用K折交叉验证，train_test_split自主切分数据集

模型选择：采用了决策树，提升树（GBDT/XGBoost），SVM（libsvm）神经网络（多层感知器算法）分别训练模型

3：对应python主要代码：

• decisiontree.py

from sklearn.metrics import accuracy_score,precision_score,recall_score,f1_score

X_train,X_test,y_train,y_test=train_test_split(StS,y,test_size=0.4,random_state=0)
clf = tree.DecisionTreeClassifier()
clf = clf.fit(X_train, y_train)
pre_labels = clf.predict(X_test)
print(‘accuracy score:‘,accuracy_score(y_test,pre_labels,normalize=True))
print(‘recall score:‘,recall_score(y_test,pre_labels))
print(‘precision score:‘,precision_score(y_test,pre_labels))
print(‘f1  score:‘,f1_score(y_test,pre_labels))

• XGBoost.py

import xgboost as xgb
from sklearn.preprocessing import StandardScaler
#记录程序运行时间
import time
start_time = time.time()
from xgboost.sklearn import XGBClassifier
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score,precision_score,recall_score,f1_score,classification_report,roc_auc_score
bankChurn = pd.read_csv(‘D:/work/lost data and dictionary/test/bankChurn.csv‘)#原始数据
bankChurn_data = pd.read_csv(‘D:/work/lost data and dictionary/test/bankChurn_data.csv‘)#预处理数据
Y_train=bankChurn[‘CHUR0_CUST_I0D‘]#标签
StS=StandardScaler().fit_transform(bankChurn_data)
X_train,X_test,y_train,y_test=train_test_split(StS,Y_train,test_size=0.4,random_state=None)
print(X_train.shape, X_test.shape)
#模型参数设置
xlf = xgb.XGBClassifier(max_depth=10,
                        learning_rate=0.1,
                        n_estimators=10,
                        silent=True,
                        objective=‘binary:logistic‘,
                        nthread=-1,
                        gamma=0,
                        min_child_weight=1,
                        max_delta_step=0,
                        subsample=0.85,
                        colsample_bytree=0.7,
                        colsample_bylevel=1,
                        reg_alpha=0,
                        reg_lambda=1,
                        scale_pos_weight=1,#这个值是因为类别十分不平衡。
                        seed=1440)

xlf.fit(X_train, y_train, eval_metric=‘error‘, verbose = True, eval_set = [(X_test, y_test)],early_stopping_rounds=100)
# 计算 auc 分数、预测
preds = xlf.predict(X_test)
pre_pro = xlf.predict_proba(X_test)[:,1]
print(‘accuracy score:‘,accuracy_score(y_test,preds ,normalize=True))
print(‘classification report:‘,classification_report(y_test,preds ))
print(‘precision score:‘,precision_score(y_test,preds ))
print(‘roc_auc_score:%f‘ % roc_auc_score(y_test,pre_pro))
#输出运行时长
cost_time = time.time()-start_time
print("xgboost success!",‘\n‘,"cost time:",cost_time,"(s)......")

• libsvm.py

import os
os.chdir(‘C:\libsvm-2.81\python‘)
from svmutil import *
from sklearn.metrics import accuracy_score,classification_report
y,x=svm_read_problem(‘bankchurnLibsvm.txt‘)#转换成libsvm格式
# print(type(x))
x=np.array(x)
y=np.array(y)
stratified_folder=StratifiedKFold(n_splits=4,random_state=0,shuffle=True)
for train_index,test_index in stratified_folder.split(x,y):
    print(‘shuffled train index:‘,train_index)
    print(‘shuffled test index:‘, test_index)
    print(‘shuffled x_train:‘, x[train_index])
    print(‘shuffled x_test:‘, x[test_index])
    print(‘shuffled y_train:‘, y[train_index])
    print(‘shuffled y_test:‘, y[test_index])
    print(‘.......‘)
y_train=list(y[train_index])
y_test=list(y[test_index])
x_train=list(x[train_index])
x_test=list(x[test_index])
m=svm_train( y_train,x_train,‘-c 4  -g 2‘)
p_label,p_acc,p_val=svm_predict(y_test,x_test,m)
print(‘accuracy score:‘,accuracy_score(y_test,p_label ,normalize=True))
print(‘classification report:‘,classification_report(y_test,p_label ))

• BPtest

  import pandas as pd
  import numpy as np
  from sklearn.model_selection import cross_val_score
  from sklearn.neural_network import MLPClassifier
  from sklearn.metrics import accuracy_score,roc_auc_score
  from sklearn.metrics import accuracy_score,precision_score,recall_score,f1_score,classification_report
  bankChurn = pd.read_csv(‘D:/work/lost data and dictionary/test/bankChurn.csv‘)
  X_data = pd.read_csv(‘D:/work/lost data and dictionary/test/bankChurn_data.csv‘)
  X_data=X_data.values[:,:]
  Y_label=bankChurn[‘CHUR0_CUST_I0D‘]
  Y_label=Y_label.values[:]
  data=np.hstack((X_data,Y_label.reshape(Y_label.size,1)))##将样本集与标签合并
  np.random.shuffle(data)##混洗数据
  X=data[:,:-1]
  Y=data[:,-1]
  train_x=X[:-8620]
  test_x=X[-8620:]
  train_y=Y[:-8620]
  test_y=Y[-8620:]#数据5:5
  ######mlpclassifier_data():###多层感知机算法，BP算法
  classifier=MLPClassifier(hidden_layer_sizes=(30,),activation=‘logistic‘,max_iter=1000)
  clf=classifier.fit(train_x,train_y)
  train_score=classifier.score(train_x,train_y)
  test_score=classifier.score(test_x,test_y)
  print(‘train_score:‘,train_score)
  print(‘test_score:‘,test_score)
  ####得到其他分类效果####
  pre_labels = clf.predict(test_x)
  pre_pro = clf.predict_proba(test_x)[:,1]
  print(‘accuracy score:‘,accuracy_score(test_y,pre_labels,normalize=True))
  print(‘recall score:‘,recall_score(test_y,pre_labels))
  print(‘classification report:‘,classification_report(test_y,pre_labels))
  print(‘precision score:‘,precision_score(test_y,pre_labels))
  print(‘f1  score:‘,f1_score(test_y,pre_labels))
  print(‘roc_auc_score:%f‘ % roc_auc_score(test_y,pre_pro))
  

  运行结果比较：

  	DT	XGBoost	Libsvm	BP
  Accuracy	0.856	0.91	0.894	0.90
  Precision	0.86	0.89	0.84	0.88
  Recall	0.86	0.91	0.89	0.90
  F1 score	0.86	0.89	0.85	0.87