Python:机器学习的ROC曲线、AUC计算，以及相应的代码

ROC曲线的原理以及绘制方法参考点击打开链接，这里主要是对原理部分的代码实现。

对于每一个给定的阈值threshold，我们都可以算出有关的TPR、FPR参数，这里我写了以下函数来实现该功能，函数的输入有result和thres两部分。前一部分是包含两个array，第一个array用来存储每一个样本是正样本概率，第二个array则是每个样本的label属性(0或1)；后一部分则是选取的阈值，代码实现原理同参考文献中相同：

def cal_rate(result, thres):
    all_number = len(result[0])
    # print all_number
    TP = 0
    FP = 0
    FN = 0
    TN = 0
    for item in range(all_number):
        disease = result[0][item]
        if disease >= thres:
            disease = 1
        if disease == 1:
            if result[1][item] == 1:
                TP += 1
            else:
                FP += 1
        else:
            if result[1][item] == 0:
                TN += 1
            else:
                FN += 1
    # print TP+FP+TN+FN
    accracy = float(TP+FP) / float(all_number)
    if TP+FP == 0:
        precision = 0
    else:
        precision = float(TP) / float(TP+FP)
    TPR = float(TP) / float(TP+FN)
    TNR = float(TN) / float(FP+TN)
    FNR = float(FN) / float(TP+FN)
    FPR = float(FP) / float(FP+TN)
    # print accracy, precision, TPR, TNR, FNR, FPR
    return accracy, precision, TPR, TNR, FNR, FPR

这只是对一个阈值进行的计算，要想设置连续的阈值计算，则需要对样本正确率进行一个升序遍历取值当阈值就可以了：

#prob是样本正确率的array，label则是样本label的array
threshold_vaule = sorted(prob)
threshold_num = len(threshold_vaule)
accracy_array = np.zeros(threshold_num)
precision_array = np.zeros(threshold_num)
TPR_array = np.zeros(threshold_num)
TNR_array = np.zeros(threshold_num)
FNR_array = np.zeros(threshold_num)
FPR_array = np.zeros(threshold_num)
# calculate all the rates
for thres in range(threshold_num):
    accracy, precision, TPR, TNR, FNR, FPR = cal_rate((prob,label), threshold_vaule[thres])
    accracy_array[thres] = accracy
    precision_array[thres] = precision
    TPR_array[thres] = TPR
    TNR_array[thres] = TNR
    FNR_array[thres] = FNR
    FPR_array[thres] = FPR

最后，利用计算公式根画图函数可以画出ROC曲线以及我们用来评估模型的AUC和ERR参数。

AUC = np.trapz(TPR_array, FPR_array)
threshold = np.argmin(abs(FNR_array - FPR_array))
EER = (FNR_array[threshold]+FPR_array[threshold])/2
plt.plot(FPR_array, TPR_array)
plt.title('roc')
plt.xlabel('FPR_array')
plt.ylabel('TPR_array')
plt.show()

最后的最后，附上总的代码。总的代码和上面几个有所不同，是我用来处理医疗数据多分类标签的问题，可忽略～

import numpy as np
import cv2
import os
import matplotlib.pyplot as plt
import re

'''
calculate each rate
'''
def cal_rate(result, num, thres):
	all_number = len(result[0])
	# print all_number
	TP = 0
	FP = 0
	FN = 0
	TN = 0
	for item in range(all_number):
		disease = result[0][item,num]
		if disease >= thres:
			disease = 1
		if disease == 1:
			if result[1][item,num] == 1:
				TP += 1
			else:
				FP += 1
		else:
			if result[1][item,num] == 0:
				TN += 1
			else:
				FN += 1
	# print TP+FP+TN+FN
	accracy = float(TP+FP) / float(all_number)
	if TP+FP == 0:
		precision = 0
	else:
		precision = float(TP) / float(TP+FP)
	TPR = float(TP) / float(TP+FN)
	TNR = float(TN) / float(FP+TN)
	FNR = float(FN) / float(TP+FN)
	FPR = float(FP) / float(FP+TN)
	# print accracy, precision, TPR, TNR, FNR, FPR
	return accracy, precision, TPR, TNR, FNR, FPR

disease_class = ['Atelectasis','Cardiomegaly','Effusion','Infiltration','Mass','Nodule','Pneumonia','Pneumothorax']
style = ['r-','g-','b-','y-','r--','g--','b--','y--']
'''
plot roc and calculate AUC/ERR, result: (prob, label) 
'''
prob = np.random.rand(100,8)
label = np.where(prob>=0.5,prob,0)
label = np.where(label<0.5,label,1)
count = np.count_nonzero(label)
label = np.zeros((100,8))
label[1:20,:]=1
print label
print prob
print count
for clss in range(len(disease_class)):
	threshold_vaule = sorted(prob[:,clss])
	threshold_num = len(threshold_vaule)
	accracy_array = np.zeros(threshold_num)
	precision_array = np.zeros(threshold_num)
	TPR_array = np.zeros(threshold_num)
	TNR_array = np.zeros(threshold_num)
	FNR_array = np.zeros(threshold_num)
	FPR_array = np.zeros(threshold_num)
	# calculate all the rates
	for thres in range(threshold_num):
		accracy, precision, TPR, TNR, FNR, FPR = cal_rate((prob,label), clss, threshold_vaule[thres])
		accracy_array[thres] = accracy
		precision_array[thres] = precision
		TPR_array[thres] = TPR
		TNR_array[thres] = TNR
		FNR_array[thres] = FNR
		FPR_array[thres] = FPR
	# print TPR_array
	# print FPR_array
	AUC = np.trapz(TPR_array, FPR_array)
	threshold = np.argmin(abs(FNR_array - FPR_array))
	EER = (FNR_array[threshold]+FPR_array[threshold])/2
	print ('disease %10s threshold : %f' % (disease_class[clss],threshold))
	print ('disease %10s accracy : %f' % (disease_class[clss],accracy_array[threshold]))
	print ('disease %10s EER : %f AUC : %f' % (disease_class[clss],EER, -AUC))
	plt.plot(FPR_array, TPR_array, style[clss], label=disease_class[clss])
plt.title('roc')
plt.xlabel('FPR_array')
plt.ylabel('TPR_array')
plt.legend()
plt.show()