评估指标:ROC,AUC,Precision、Recall、F1-score

一、ROC,AUC

ROC(Receiver Operating Characteristic)曲线和AUC常被用来评价一个二值分类器(binary classifier)的优劣 。

ROC曲线一般的横轴是FPR,纵轴是FPR。AUC为曲线下面的面积,作为评估指标,AUC值越大,说明模型越好。如下图:

 

二、Precision、Recall、F1-score

Terminology and derivations
from a confusion matrix
true positive (TP)
eqv. with hit
true negative (TN)
eqv. with correct rejection
false positive (FP)
eqv. with false alarmType I error
false negative (FN)
eqv. with miss, Type II error

sensitivity or true positive rate (TPR)
eqv. with hit raterecall
\mathit{TPR} = \frac {\mathit{TP}} {P} = \frac {\mathit{TP}} {\mathit{TP}+\mathit{FN}}
specificity (SPC) or true negative rate (TNR)
\mathit{SPC} = \frac {\mathit{TN}} {N} = \frac {\mathit{TN}} {\mathit{FP} + \mathit{TN}}
precision or positive predictive value (PPV)
\mathit{PPV} = \frac {\mathit{TP}} {\mathit{TP} + \mathit{FP}}
negative predictive value (NPV)
\mathit{NPV} = \frac {\mathit{TN}} {\mathit{TN} + \mathit{FN}}
fall-out or false positive rate (FPR)
\mathit{FPR} = \frac {\mathit{FP}} {N} = \frac {\mathit{FP}} {\mathit{FP} + \mathit{TN}} = 1 - \mathit{SPC}
false discovery rate (FDR)
\mathit{FDR} = \frac {\mathit{FP}} {\mathit{FP} + \mathit{TP}} = 1 - \mathit{PPV}
miss rate or false negative rate (FNR)
\mathit{FNR} = \frac {\mathit{FN}} {P} = \frac {\mathit{FN}} {\mathit{FN} + \mathit{TP}}

accuracy (ACC)
\mathit{ACC} = \frac {\mathit{TP} + \mathit{TN}} {P + N}
F1 score
is the harmonic mean of precision and sensitivity
\mathit{F1} = \frac {2 \mathit{TP}} {2 \mathit{TP} + \mathit{FP} + \mathit{FN}}
Matthews correlation coefficient (MCC)
\frac{ TP \times TN - FP \times FN } {\sqrt{ (TP+FP) ( TP + FN ) ( TN + FP ) ( TN + FN ) } }

Informedness = Sensitivity + Specificity - 1
Markedness = Precision + NPV - 1

中文解释

posted @ 2018-08-04 22:35  杨国峰  阅读(729)  评论(0编辑  收藏  举报