Estimate the area under the ROC curve from positive and negative prediction scores, to compare the ranking ability of classifiers.
基于正态分布假设的近似。
| Item | Formula / Note | Source / Standard |
|---|---|---|
| AUC meaning | the area equals the probability a random positive outranks a random negative | Fawcett 2006, Pattern Recognition Letters |
| Trapezoidal estimate | sorting by score and summing trapezoids approximates the area under the curve | Hanley & McNeil 1982, Radiology |
| Threshold independence | the metric is threshold independent which suits comparing ranking power across models | Metz 1978, Seminars in Nuclear Medicine |
AUC 为 ROC 曲线下面积,等价于随机正样本得分高于随机负样本的概率。