For one of my binary classification models, I have observed this (Simpson's Rule-esque) paradox. The AUC on the test set as a whole is 0.8.

Gender is one of the model's features. So I decided to produce a "bias" report, for which I calculated AUCs for each of the Male and Female subgroups. But I noticed that each of these AUCs is around 0.7. How is this possible, given that the overall test AUC is 0.8? (In my dataset, every data point belongs to either the Male or the Female subgroup.) I don't expect the overall AUC to simply be a (weighted) linear combination of AUCs for the individual strata.

I'm hoping to get both a technical/mathematical answer and a high-level explanation. Please let me know if any further information is needed (if you think I should plot the overall, Male, and Female ROC curves, for instance). Thank you!

  • 1
    $\begingroup$ From @Valentas, take an extreme example. All y=0, males score < 0.25, y=1 males (0.26-0.5), y=0 females (0.51-0.75) and y=1 females (0.76-1). When looking at just males or females the AUROC is 1 since y=0 scores < y=1 scores. Now think what happens when combing the data. Male y=1 scores are < female y=0 scores. This is one of the weaknesses of AUROC. $\endgroup$
    – Craig
    Mar 19 at 9:52
  • $\begingroup$ Thanks, @Craig! But how would you modify this example to show the opposite: a high overall AUC but relatively low AUCs for each of the Male and Female subgroups? Appreciate your help. $\endgroup$ Mar 20 at 1:58
  • $\begingroup$ Plot your data. Scores on the X-axis. Males y=0 and males y=1 as separate colors and/or shapes. Now you can see what is rank-ordered correctly vs incorrectly. Again with females on a separate plot. Now combined. You can visually see. Sample and jitter as needed. Make sure AUROC is the right metric to use for your business problem. I am not always interested in rank-order power so AUROC is not always applicable. To understand more about AUROC - stats.stackexchange.com/questions/145566/…). $\endgroup$
    – Craig
    Mar 21 at 10:43

AUC can be defined as $P(X_1 > X_0)$ where $X_1$ is the score of a randomly chosen positive instance and $X_0$ is the score of a randomly chosen negative instance.

Like in Simpson's "paradox", what you see could happen because the group has a relatively large effect on your target. For example, one group could be mostly positive, another could be mostly negative. The classifier might not work at all within each of the groups, but if it is merely able to predict the group we are in (or as in your case it is simply one of the inputs), it can have a high AUC.


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.