python_数据_roc_auc

auc roc

start

import numpy as np 
import pandas as pd
import matplotlib.pyplot as plt

from sklearn import datasets

from sklearn.metrics import auc
from sklearn.metrics import roc_curve
from sklearn.metrics import confusion_matrix
from sklearn.metrics import classification_report

from sklearn.model_selection import StratifiedKFold
from sklearn.model_selection import train_test_split

from sklearn.svm import SVC
from sklearn.linear_model import LogisticRegression

假设有一组数据,prob(特征)与y(分类)具有关系
使用不同的阈值(threshold),通过阈值对prob进行划分

y = np.array([1,1,1,1,1,1,0,0,0,1,1,0,0,1,1,0,0,1,1,0,0])
prob = np.array([0.42,0.73,0.55,0.37,0.57,0.70,0.25,0.23,0.46,0.62,0.76,0.46,0.55,0.56,0.56,0.38,0.37,0.73,0.77,0.21,0.39])

thresholds = np.linspace(0,1,11)

for threshold in thresholds:
    y_ = (prob >= threshold).astype(np.int8)
    print(threshold.astype(np.float16),end='\t')
    print((y_ == y).mean().astype(np.float16))

• out: 0.0 0.5713
  0.1 0.5713
  0.2 0.5713
  0.3 0.7144
  0.4 0.8096
  0.5 0.857
  0.6 0.7144
  0.7 0.619
  0.8 0.4285
  0.9 0.4285
  1.0 0.4285
• 此时,当阈值为0.5时,准确率最高

选取的阈值为0.4时,观察交叉表

y_ = (prob >= 0.4).astype(np.int8)

cm = pd.crosstab(y_,y,rownames=['预测'],colnames=['真实'])
cm

• out:

预测 \真实	0	1
0	6	1
1	3	11

# 交叉表的数据
cm = confusion_matrix(y_,y,)
cm[0,0],cm

• out: (6, array([[ 6, 1],
  [ 3, 11]], dtype=int64))

绘制roc曲线与洛伦兹曲线

tprs = []
fprs = []
for threshold in thresholds:
    # 给了阈值,概率变成了类别
    y_ = (prob >= threshold).astype(np.int8)
    cm = confusion_matrix(y_,y)
    
    tpr = cm[1,1]/(cm[0,1] + cm[1,1])   # 召回率 越大越好
    fpr = cm[1,0]/(cm[0,0] + cm[1,0])   # 越小越好
    
    tprs.append(tpr)
    fprs.append(fpr)
    
plt.plot(fprs,tprs)  # x轴为fprs,y轴为召回率tprs;曲线右下方空间越大越好
auc(fprs,tprs)   # 比准确率更有说服性   0.9212962962962963

y_1 = np.array(tprs) - np.array(fprs)   # 作为y轴
x_1 = np.linspace(0,1,11)   # 作为x轴
# ??画在同一图中,坐标意义不同
plt.plot(x_1,y_1)    # 洛伦兹曲线   # 正负样本分离度

• out:
  

使用datasets中的酒数据,使用svm(支持向量机)和逻辑斯蒂两种方式对比分类
对分类结果进行列表展示

wine = datasets.load_wine()
X = wine['data']
y = wine['target']

wine   # 略
wine.target_names  # 三分类问题

• out: array([‘class_0’, ‘class_1’, ‘class_2’], dtype=’<U7’)

svc = SVC(kernel='poly',degree=3,gamma='auto')

X_train,X_test,y_train,y_test = train_test_split(X,y,random_state = 2)
svc.fit(X_train,y_train)
y_ = svc.predict(X_test)

print(classification_report(y_test,y_,labels=[0,1,2]))
print(svc.score(X_test,y_test))

• out:

•          		precision    recall  f1-score   support
  
           0       1.00      1.00      1.00        19
           1       0.93      1.00      0.96        13
           2       1.00      0.92      0.96        13
   micro avg       0.98      0.98      0.98        45
   macro avg       0.98      0.97      0.97        45
  weighted avg     0.98      0.98      0.98        45
  

• 0.9777777777777777

logis = LogisticRegression()

X_train,X_test,y_train,y_test = train_test_split(X,y,random_state = 2)
logis.fit(X_train,y_train)
y_ = logis.predict(X_test)

print(classification_report(y_test,y_,labels=[0,1,2]))
print(logis.score(X_test,y_test))

• out :

•               precision    recall  f1-score   support
  
           0       1.00      0.89      0.94        19
           1       0.81      1.00      0.90        13
           2       1.00      0.92      0.96        13
   micro avg       0.93      0.93      0.93        45
   macro avg       0.94      0.94      0.93        45
  weighted avg     0.95      0.93      0.94        45
  

• 0.9333333333333333

支持向量机和逻辑斯蒂都更倾向于二分类
样本不均衡或二分类时,更要用auc
抽取酒数据的的两类作为数据,绘制auc曲线

cond = (y != 2)
X = X[cond]
y = y[cond]

model = SVC(probability=True,gamma='auto',kernel='poly')

# 六折交叉验证
sKFold = StratifiedKFold(n_splits=6)

i = 0
for train,test in sKFold.split(X,y):
    X_train,X_test = X[train],X[test]
    y_train,y_test = y[train],y[test]
    
    model.fit(X_train,y_train)

    proba_ = model.predict_proba(X_test)

    y_ = proba_[:,1]  # y_为预测其为正样本(1)的概率

#     fpr,tpr,thresholds = roc_curve(y_test,y_,drop_intermediate=False)
    
    fpr,tpr,thresholds = roc_curve(y_test,y_)
    
    auc_ = auc(fpr,tpr)

    plt.plot(fpr,tpr,label = 'Fold {} AUC {}'.format(i,auc_))

    i+=1
plt.legend()

• out:
  

model = LogisticRegression(solver='liblinear')

sKFold = StratifiedKFold(n_splits=6)

i = 0
for train,test in sKFold.split(X,y):
    X_train,X_test = X[train],X[test]
    y_train,y_test = y[train],y[test]
    
    model.fit(X_train,y_train)

    proba_ = model.predict_proba(X_test)

    y_ = proba_[:,1]

    fpr,tpr,thresholds = roc_curve(y_test,y_)

    auc_ = auc(fpr,tpr)     # 可以用平均auc

    plt.plot(fpr,tpr,label = 'Fold {} AUC {}'.format(i,auc_))

    i+=1
plt.legend()

• out:
  

• fpr,tpr,thresholds = roc_curve(y_test,y_)

  参数

  y_true : array, shape = [n_samples]
  True binary labels. If labels are not either {-1, 1} or {0, 1}, then
  pos_label should be explicitly given.

    y_test
  

  y_score : array, shape = [n_samples]
  Target scores, can either be probability estimates of the positive
  class, confidence values, or non-thresholded measure of decisions
  (as returned by “decision_function” on some classifiers).
  pos_label : int or str, default=None
  Label considered as positive and others are considered negative.

    y_ -----> probability estimates of the positive class
  

  drop_intermediate : boolean, optional (default=True)
  Whether to drop some suboptimal thresholds which would not appear
  on a plotted ROC curve. This is useful in order to create lighter
  ROC curves.

  返回值

  fpr : array, shape = [>2]
  Increasing false positive rates such that element i is the false
  positive rate of predictions with score >= thresholds[i].

  tpr : array, shape = [>2]
  Increasing true positive rates such that element i is the true
  positive rate of predictions with score >= thresholds[i].

  thresholds : array, shape = [n_thresholds]
  Decreasing thresholds on the decision function used to compute
  fpr and tpr. thresholds[0] represents no instances being predicted
  and is arbitrarily set to max(y_score) + 1.