構建７種分類模型，評分並畫出ROC曲線

• 匯入的包

import pandas as pd
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression
from sklearn.tree import DecisionTreeClassifier
from sklearn.svm import SVC
from sklearn.metrics import f1_score,precision_score,recall_score,roc_auc_score,accuracy_score,roc_curve
import matplotlib.pyplot as plt
from sklearn.ensemble import RandomForestClassifier
from sklearn.ensemble import GradientBoostingClassifier
from xgboost.sklearn import XGBClassifier
import lightgbm as lgb
 

• 讀取資料集

data_all = pd.read_csv('/home/infisa/wjht/project/DataWhale/data_all.csv', encoding='gbk')

• 劃分訓練集和測試集

features = [x for x in data_all.columns if x not in ['status']]
X = data_all[features]
y = data_all['status']
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=2018)
 

• 構建模型

lr = LogisticRegression(random_state=2018,tol=1e-6)  # 邏輯迴歸模型
lr.fit(X_train, y_train)
lr_y_proba=lr.predict_proba(X_test)
lr_y_pre=lr.predict(X_test)

tr = DecisionTreeClassifier(random_state=2018)  # 決策樹模型
tr.fit(X_train, y_train)
tr_y_pre=tr.predict(X_test)
tr_y_proba=tr.predict_proba(X_test)

svm = SVC(probability=True,random_state=2018,tol=1e-6)  # SVM模型
svm.fit(X_train, y_train)
svm_y_pre=svm.predict(X_test)
svm_y_proba=svm.predict_proba(X_test)

forest=RandomForestClassifier(n_estimators=100,random_state=2018) #　隨機森林
forest.fit(X_train,y_train)
forest_y_pre=forest.predict(X_test)
forest_y_proba=forest.predict_proba(X_test)



Gbdt=GradientBoostingClassifier(random_state=2018) #CBDT
Gbdt.fit(X_train,y_train)
Gbdt_y_pre=Gbdt.predict(X_test)
Gbdt_y_proba=Gbdt.predict_proba(X_test)


Xgbc=XGBClassifier(random_state=2018)  #Xgbc
Xgbc.fit(X_train,y_train)
Xgbc_y_pre=Xgbc.predict(X_test)


gbm=lgb.LGBMClassifier(random_state=2018)  #lgb
gbm.fit(X_train,y_train)
gbm_y_pre=gbm.predict(X_test)
gbm_y_proba=gbm.predict_proba(X_test)
 

• 模型評分

# 模型評分
lr_score = lr.score(X_test, y_test)
lr_accuracy_score=accuracy_score(y_test,lr_y_pre)
lr_preci_score=precision_score(y_test,lr_y_pre)
lr_recall_score=recall_score(y_test,lr_y_pre)
lr_f1_score=f1_score(y_test,lr_y_pre)
lr_auc=roc_auc_score(y_test,lr_y_proba[:,1])
print('lr_accuracy_score: %f,lr_preci_score: %f,lr_recall_score: %f,lr_f1_score: %f,lr_auc: %f'
      %(lr_accuracy_score,lr_preci_score,lr_recall_score,lr_f1_score,lr_auc))
'lr_accuracy_score: 0.768746,lr_preci_score: 0.688312,lr_recall_score: 0.147632,lr_f1_score: 0.243119,lr_auc: 0.716681'


tr_score = tr.score(X_test, y_test)
tr_accuracy_score=accuracy_score(y_test,tr_y_pre)
tr_preci_score=precision_score(y_test,tr_y_pre)
tr_recall_score=recall_score(y_test,tr_y_pre)
tr_f1_score=f1_score(y_test,tr_y_pre)
tr_auc=roc_auc_score(y_test,tr_y_proba[:,1])
# print('tr_accuracy_score: %f,tr_preci_score: %f,tr_recall_score: %f,tr_f1_score: %f,tr_auc: %f'
#       %(tr_accuracy_score,tr_preci_score,tr_recall_score,tr_f1_score,tr_auc))

'tr_accuracy_score: 0.684653,tr_preci_score: 0.382429,tr_recall_score: 0.412256,tr_f1_score: 0.396783,tr_auc: 0.594237'



svm_accuracy_score=accuracy_score(y_test,svm_y_pre)
svm_preci_score=precision_score(y_test,svm_y_pre)
svm_recall_score=recall_score(y_test,svm_y_pre)
svm_f1_score=f1_score(y_test,svm_y_pre)
svm_auc=roc_auc_score(y_test,svm_y_proba[:,1])
print('svm_accuracy_score: %f,svm_preci_score: %f,svm_recall_score: %f,svm_f1_score: %f,svm_auc: %f'
      %(svm_accuracy_score,svm_preci_score,svm_recall_score,svm_f1_score,svm_auc))
'svm_accuracy_score: 0.748423,svm_preci_score: 0.000000,svm_recall_score: 0.000000,svm_f1_score: 0.000000,svm_auc: 0.500000'




forest_accuracy_score=accuracy_score(y_test,forest_y_pre)
forest_preci_score=precision_score(y_test,forest_y_pre)
forest_recall_score=recall_score(y_test,forest_y_pre)
forest_f1_score=f1_score(y_test,forest_y_pre)
forest_auc=roc_auc_score(y_test,forest_y_proba[:,1])
print('forest_accuracy_score: %f,forest_preci_score: %f,forest_recall_score: %f,forest_f1_score: %f,forest_auc: %f'
      %(forest_accuracy_score,forest_preci_score,forest_recall_score,forest_f1_score,forest_auc))
'forest_accuracy_score: 0.782060,forest_preci_score: 0.681818,forest_recall_score: 0.250696,forest_f1_score: 0.366599,forest_auc: 0.749137'



Gbdt_accuracy_score=accuracy_score(y_test,Gbdt_y_pre)
Gbdt_preci_score=precision_score(y_test,Gbdt_y_pre)
Gbdt_recall_score=recall_score(y_test,Gbdt_y_pre)
Gbdt_f1_score=f1_score(y_test,Gbdt_y_pre)
Gbdt_auc=roc_auc_score(y_test,Gbdt_y_proba[:,1])
print('Gbdt_accuracy_score: %f,Gbdt_preci_score: %f,Gbdt_recall_score: %f,Gbdt_f1_score: %f,Gbdt_auc: %f'
      %(Gbdt_accuracy_score,Gbdt_preci_score,Gbdt_recall_score,Gbdt_f1_score,Gbdt_auc))

'Gbdt_accuracy_score: 0.780659,Gbdt_preci_score: 0.611650,Gbdt_recall_score: 0.350975,Gbdt_f1_score: 0.446018,Gbdt_auc: 0.763828'


Xgbc_accuracy_score=accuracy_score(y_test,Xgbc_y_pre)
Xgbc_preci_score=precision_score(y_test,Xgbc_y_pre)
Xgbc_recall_score=recall_score(y_test,Xgbc_y_pre)
Xgbc_f1_score=f1_score(y_test,Xgbc_y_pre)
Xgbc_auc=roc_auc_score(y_test,Xgbc_y_pre)
print('Xgbc_accuracy_score: %f,Xgbc_preci_score: %f,Xgbc_recall_score: %f,Xgbc_f1_score: %f,Xgbc_auc: %f'
      %(Xgbc_accuracy_score,Xgbc_preci_score,Xgbc_recall_score,Xgbc_f1_score,Xgbc_auc))
'Xgbc_accuracy_score: 0.785564,Xgbc_preci_score: 0.630542,Xgbc_recall_score: 0.356546,Xgbc_f1_score: 0.455516,Xgbc_auc: 0.643161'



gbm_accuracy_score=accuracy_score(y_test,gbm_y_pre)
gbm_preci_score=precision_score(y_test,gbm_y_pre)
gbm_recall_score=recall_score(y_test,gbm_y_pre)
gbm_f1_score=f1_score(y_test,gbm_y_pre)
gbm_auc=roc_auc_score(y_test,gbm_y_proba[:,1])
print('gbm_accuracy_score: %f,gbm_preci_score: %f,gbm_recall_score: %f,gbm_f1_score: %f,gbm_auc: %f'
      %(gbm_accuracy_score,gbm_preci_score,gbm_recall_score,gbm_f1_score,gbm_auc))
'gbm_accuracy_score: 0.770147,gbm_preci_score: 0.570136,gbm_recall_score: 0.350975,gbm_f1_score: 0.434483,gbm_auc: 0.757402'

• 畫出Roc曲線

lr_fpr,lr_tpr,lr_threasholds=roc_curve(y_test,lr_y_proba[:,1]) # 計算ROC的值,lr_threasholds為閾值
plt.title("roc_curve of %s(AUC=%.4f)" %('logist',lr_auc))
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.plot(lr_fpr,lr_tpr)
plt.show()

tr_fpr,tr_tpr,tr_threasholds=roc_curve(y_test,tr_y_proba[:,1]) # 計算ROC的值,lr_threasholds為閾值
plt.title("roc_curve of %s(AUC=%.4f)" %('decisiontree',tr_auc))
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.plot(tr_fpr,tr_tpr)
plt.show()

svm_fpr,svm_tpr,svm_threasholds=roc_curve(y_test,svm_y_proba[:,1]) # 計算ROC的值,svm_threasholds為閾值
plt.title("roc_curve of %s(AUC=%.4f)" %('svm',svm_auc))
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.plot(svm_fpr,svm_tpr)
plt.show()

forest_fpr,forest_tpr,forest_threasholds=roc_curve(y_test,forest_y_proba[:,1]) # 計算ROC的值,svm_threasholds為閾值
plt.title("roc_curve of %s(AUC=%.4f)" %('forest',forest_auc))
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.plot(forest_fpr,forest_tpr)
plt.show()

Gbdt_fpr,Gbdt_tpr,Gbdt_threasholds=roc_curve(y_test,Gbdt_y_proba[:,1]) # 計算ROC的值,svm_threasholds為閾值
plt.title("roc_curve of %s(AUC=%.4f)" %('Gbdt',Gbdt_auc))
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.plot(Gbdt_fpr,Gbdt_tpr)
plt.show()

Xgbc_fpr,Xgbc_tpr,Xgbc_threasholds=roc_curve(y_test,Xgbc_y_pre) # 計算ROC的值,svm_threasholds為閾值
plt.title("roc_curve of %s(AUC=%.4f)" %('Xgbc',Xgbc_auc))
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.plot(Xgbc_fpr,Xgbc_tpr)
plt.show()

gbm_fpr,gbm_tpr,gbm_threasholds=roc_curve(y_test,gbm_y_proba[:,1]) # 計算ROC的值,svm_threasholds為閾值
plt.title("roc_curve of %s(AUC=%.4f)" %('gbm',gbm_auc))
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.plot(gbm_fpr,gbm_tpr)
plt.show()

• 思考
  01 對於roc曲線，直接畫成直線不太理解；
  02 對於svm_accuracy_score: 0.748423,svm_preci_score: 0.000000,svm_recall_score: 0.000000,svm_f1_score: 0.000000,svm_auc: 0.500000　其中得出的項為什麼為０不理解
• 參考的文章
  機器學習中的 precision、recall、accuracy、F1 Score
  分類問題的幾個評價指標（Precision、Recall、F1-Score、Micro-F1、Macro-F1）