一、指标概念
TP - 将正类预测为正类
FN - 将正类预测为负类, 类似假设检验中的第一类错误,拒真
FP - 将负类预测位正类,类似假设检验中的第二类错误,取伪
TN - 将负类预测位负类
假设检验第一类错误(Type I error):原假设是正确的,却拒绝了原假设。
一个例子:某池塘有1400条鲤鱼,300只虾,300只鳖
现在以捕鲤鱼为目的。那么,这些指标分别如下:
逮着了700条鲤鱼,200只虾,100只鳖 | 把池子里的所有的鲤鱼、虾和鳖都一网打尽 | |
Precision | 700 / (700 + 200 + 100) = 70% | 1400 / (1400 + 300 + 300) = 70% |
Recall | 700 / 1400 = 50% | 1400 / 1400 = 100% |
F值 | 70% * 50% * 2 / (70% + 50%) = 58.3% | 70% * 100% * 2 / (70% + 100%) = 82.35% |
另一个例子
生病检查数据样本有10000个,其中10个数据祥本是有病,其它是无病。
假设分类模型在无病数据9990中预测正确了9980个,在10个有病数据中预测正确了9个,真阳=9,真阴=9980,假阳=10,假阴=1
可翻译为 TP=9, FN(拒真)=1, FP(取伪)= 10,TN=9980
于是
Precision= TP / (TP +FP) = 9/(9+10)= 47.36%
Accuracy = (TP + TN) /(TP + FN + FP + TN)= (9+9980) /10000=99.89%
Recall = TP / (TP + FN) = 9/(9+1) = 90%
F-Measure(F值) = 2 * Recall * Precision/ (Precision+Recall)
F1-score=2 *Recall *Precision/ (Precision+Recall)= 2×(47.36% × 90%)/(1×47.36%+90%)=62.07%
F2-score=5× (47.36% × 90%)/(4×47.36%+90%)=76. 27%
参考链接:
https://blog.csdn.net/u011630575/article/details/80250177
https://blog.csdn.net/xwd18280820053/article/details/70674256
https://blog.csdn.net/saltriver/article/details/74012163