13
$\begingroup$

One common measure used to compare two or more classification models is to use the area under the ROC curve (AUC) as a way to indirectly assess their performance. In this case a model with a larger AUC is usually interpreted as performing better than a model with a smaller AUC. But , according to Vihinen, 2012 (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3303716/), when both curves cross each other such comparison is no longer valid. Why is it so?

For instance what one could ascertain about models A, B and C based on the ROC curves and AUCs bellow?

enter image description here

$\endgroup$
19
$\begingroup$

A ROC curve visualizes TPR and FPR for all possible thresholds.

  • If you plot two ROC curves 'A' and 'B' and they do not cross each other, then one of your classifiers is clearly performing better, because for all possible FPR values you get a higher TPR. Obviously the area under the ROC will also be greater.

  • Now, if they do cross each other, then there is a point where FPR and TPR are the same for both curves 'A' and 'B'. You can no longer say that one ROC curve performs better, as it now depends on what trade-off you prefer. Do you want high Precision/low Recall or low Precision/high Recall ?

Example: If one classifier performs much better on a FPR of 0.2, but it is important to reach high Recall, then it performs well on a threshold you are not interested in.

About the ROC curves in your graph: You can easily tell that 'A' performs much better, without even knowing what you want to achieve. As soon as the violet curve crosses the others it crosses them again. You are most probably not interested in that small part, where 'B' and 'C' perform slighty better.

In the following graph you see two ROC curves, that also cross each other. Here, you cannot tell which one is better as they kind of complement each other.

Crossing ROC curves

Notice, that at the end of the day, you are interested in picking one threshold for your classification and the AUC only gives you an estimation of how well a model performs in general.

$\endgroup$
  • $\begingroup$ Just for confirmation, in my example, if I choose a very high cut-off value, where precision is large, the model A would out-perform B and C by a good margin? $\endgroup$ – Edu Sep 22 '17 at 17:33
  • $\begingroup$ How do you know where precision is large? This is a ROC, not a Precision-Recall curve. For Precision you check the probability of True Positive, given your classifier said Positive. $\endgroup$ – lnathan Sep 22 '17 at 20:25

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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