机器学习实战——使用决策树进行初始信用评级

这次使用决策树对信用进行评级
数据源与上两篇博文相同
数据挖掘实例——信用评级
机器学习实例——信用评级全流程实现
详细代码与数据可以访问下面的GitHub地址
GitHub地址

读取数据

import pandas as pd
import numpy as np
churn = pd.read_csv('telecom_churn.csv')  # 读取已经整理好的数据
churn.head()

一些简单的可视化

import matplotlib.pyplot as plt
%matplotlib inline
import seaborn as sns
sns.barplot(x='edu_class', y='churn',data=churn)
plt.show()

sns.boxplot(x='churn', y='peakMinDiff', hue=None, data=churn)
plt.show()

sns.boxplot(x='churn', y='duration', hue='edu_class', data=churn)
plt.show()

筛选变量

corrmatrix = churn.corr(method='spearman')  # spearman相关系数矩阵，可选pearson相关系数，目前仅支持这两种,函数自动排除category类型
corrmatrix_new=corrmatrix[np.abs(corrmatrix) > 0.5]
churn['duration_bins'] = pd.qcut(churn.duration,5)  #  将duration字段切分为数量（大致）相等的5段
churn['churn'].astype('int64').groupby(churn['duration_bins']).agg(['count', 'mean'])
bins = [0, 4, 8, 12, 22, 73]
churn['duration_bins'] = pd.cut(churn['duration'], bins, labels=False)
churn['churn'].astype('int64').groupby(churn['duration_bins']).agg(['mean', 'count'])

选择与目标关联较大的分类变量

import sklearn.feature_selection as feature_selection

churn['gender'] = churn['gender'].astype('int')
churn['edu_class'] = churn['edu_class'].astype('int')
churn['feton'] = churn['feton'].astype('int')
feature_selection.chi2(churn[['gender', 'edu_class', 'feton', 'prom', 
                              'posPlanChange','duration_bins', 'curPlan', 'call_10086']], churn['churn'])#选取部分字段进行卡方检验
#根据结果显示，'prom'、'posPlanChange'、'curPlan'字段可以考虑排除

(array([5.52612343e+01, 1.43446961e+01, 1.19205185e+02,1.11235701e+00,
2.98137046e-03, 1.25114016e+03, 1.77204831e+00, 5.59152494e+01]),
array([1.05529433e-013, 1.52207878e-004, 9.44377274e-028, 2.91570161e-001,
9.56455603e-001, 4.69144344e-274, 1.83128787e-001, 7.56633888e-014]))

使用决策树建模

model_data = churn[['subscriberID','churn','gender','edu_class','feton','duration_bins','call_10086','AGE']]#第二可选方案
model_data.head()

import sklearn.model_selection as cross_validation
import sklearn.tree as tree
target = model_data['churn']  # 选取目标变量
data=model_data.ix[:, 'gender':]  # 选取自变量

train_data, test_data, train_target, test_target = cross_validation.train_test_split(data,target, test_size=0.4, train_size=0.6 ,random_state=12345) # 划分训练集和测试集
clf = tree.DecisionTreeClassifier(criterion='entropy', max_depth=8, min_samples_split=5) # 当前支持计算信息增益和GINI
clf.fit(train_data, train_target)  #  使用训练数据建模

# 查看模型预测结果
train_est = clf.predict(train_data)  #  用模型预测训练集的结果
train_est_p=clf.predict_proba(train_data)[:,1]  #用模型预测训练集的概率
test_est=clf.predict(test_data)  #  用模型预测测试集的结果
test_est_p=clf.predict_proba(test_data)[:,1]  #  用模型预测测试集的概率
pd.DataFrame({'test_target':test_target,'test_est':test_est,'test_est_p':test_est_p}).T # 查看测试集预测结果与真实结果对比

模型评估

import sklearn.metrics as metrics

print(metrics.confusion_matrix(test_target, test_est,labels=[0,1]))  # 混淆矩阵
print(metrics.classification_report(test_target, test_est))  # 计算评估指标
print(pd.DataFrame(list(zip(data.columns, clf.feature_importances_))))  # 变量重要性指标

[[669 112] [144 461]]
precision recall f1-score support

     0.0       0.82      0.86      0.84       781
     1.0       0.80      0.76      0.78       605


accuracy                           0.82      1386    
macro avg       0.81      0.81      0.81      1386 
weighted avg       0.81      0.82      0.81      1386  0         1 0        
           gender  0.069254 1      
           edu_class  0.074757 2         
            feton  0.059113 3  
            duration_bins  0.672283 4     
            call_10086  0.021396 5            
            AGE  0.103199

查看预测值的分布情况

red, blue = sns.color_palette("Set1", 2)
sns.distplot(test_est_p[test_target == 1], kde=False, bins=15, color=red)
sns.distplot(test_est_p[test_target == 0], kde=False, bins=15,color=blue)
plt.show()

fpr_test, tpr_test, th_test = metrics.roc_curve(test_target, test_est_p)
fpr_train, tpr_train, th_train = metrics.roc_curve(train_target, train_est_p)
plt.figure(figsize=[6,6])
plt.plot(fpr_test, tpr_test, color=blue)
plt.plot(fpr_train, tpr_train, color=red)
plt.title('ROC curve')
plt.show()

参数调优

from sklearn.model_selection import GridSearchCV
from sklearn import metrics

param_grid = {
    'criterion':['entropy','gini'],
    'max_depth':[2,3,4,5,6,7,8],
    'min_samples_split':[4,8,12,16,20,24,28] 
}
clf = tree.DecisionTreeClassifier()
clfcv = GridSearchCV(estimator=clf, param_grid=param_grid, 
                   scoring='roc_auc', cv=4)
clfcv.fit(train_data, train_target)

查看模型预测结果

train_est = clfcv.predict(train_data)  #  用模型预测训练集的结果
train_est_p=clfcv.predict_proba(train_data)[:,1]  #用模型预测训练集的概率
test_est=clfcv.predict(test_data)  #  用模型预测测试集的结果
test_est_p=clfcv.predict_proba(test_data)[:,1]  #  用模型预测测试集的概率
#%%
fpr_test, tpr_test, th_test = metrics.roc_curve(test_target, test_est_p)
fpr_train, tpr_train, th_train = metrics.roc_curve(train_target, train_est_p)
plt.figure(figsize=[6,6])
plt.plot(fpr_test, tpr_test, color=blue)
plt.plot(fpr_train, tpr_train, color=red)
plt.title('ROC curve')
plt.show()

参数调优后拟合效果好了很多，但是有一点过拟合