ROCR包中ROC曲线计算是取大于cutoff还是大于等于cutoff
找到对应的代码如下
.compute.unnormalized.roc.curve
function (predictions, labels)
{
pos.label <- levels(labels)[2]
neg.label <- levels(labels)[1]
pred.order <- order(predictions, decreasing = TRUE)
predictions.sorted <- predictions[pred.order]
tp <- cumsum(labels[pred.order] == pos.label)
fp <- cumsum(labels[pred.order] == neg.label)
dups <- rev(duplicated(rev(predictions.sorted)))
tp <- c(0, tp[!dups])
fp <- c(0, fp[!dups])
cutoffs <- c(Inf, predictions.sorted[!dups])
return(list(cutoffs = cutoffs, fp = fp, tp = tp))
}
可以看到先按从大到小排序,再累计当前位置和之前位置的阳性值。因此计算TP,TN等指标时,取的是大于等于cutoff