5
$\begingroup$

I found the ROC explanation at this link. It states that the ROC curve is TP vs FP.

After the score has gone below 0.5, all predictions are negative. That makes them either TN or FN. Thus, how does it make sense to continue drawing as the axes are stated as TP and FP? I am really confused.

Improve this question
$\endgroup$
2
  • $\begingroup$ I think this is a situation where you're probably going to get better explanations for why your understanding is wrong the better you explain what your understanding is. $\endgroup$
    – Acccumulation
    Commented Aug 12, 2021 at 3:42
  • $\begingroup$ I found this very well explained video. I suppose I got it wrong because of the lack of background theory? $\endgroup$
    – CaTx
    Commented Aug 12, 2021 at 6:12

2 Answers 2

Reset to default
3
$\begingroup$

The ROC curve shows the TPR and FPR as you change the threshold. Your question posits that a classifier only has one TPR statistic and one FPR statistic, and that both of these statistics correspond to a threshold at 0.5. This is a common misconception; there is no reason that the threshold must be 0.5. For each choice of threshold, there is a corresponding TPR and FPR statistic. The purpose of the ROC curve is to show the trade-off for each choice of threshold: as your TPR increases, so does your FPR, and vice-versa.

Improve this answer
$\endgroup$
1
  • $\begingroup$ Goodness! You caught the cause of my misunderstanding. $\endgroup$
    – CaTx
    Commented Aug 11, 2021 at 17:56
5
$\begingroup$

you vary the threshold which is used to determine positives and negatives, ie at threshold 0.5 we get a point (TP, FP) which we plot, then at threshold 0.45 we get another point (TP, FP) which we plot, etc. These points make the ROC curve

Improve this answer
$\endgroup$
4
  • $\begingroup$ I dont quite follow your explanation. Say at point 1 (0.9) in that link, which (TP, FP) is acquired from that threshold value? $\endgroup$
    – CaTx
    Commented Aug 11, 2021 at 17:58
  • 2
    $\begingroup$ @CaTx The threshold is not graphed in the ROC curve. $\endgroup$
    – Dave
    Commented Aug 11, 2021 at 20:17
  • $\begingroup$ Thank you for the clarification! $\endgroup$
    – CaTx
    Commented Aug 12, 2021 at 6:11
  • 1
    $\begingroup$ some ROCs feature the threshold e.g. as color. E.g. R package ROCR offers this option (and I recommend to use it, this additional information is often highly instructive). $\endgroup$
    – cbeleites
    Commented Aug 12, 2021 at 10:09

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.