Timeline for Why people use precision and not true negative rate (specificity)?
Current License: CC BY-SA 4.0
7 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Jan 9, 2022 at 11:49 | comment | added | Erwan | @AmitKeinan I think that this is an interesting point, but it's broader than the question of evaluation measures: the main assumption in classification is that there is some "true class distribution" that both the training and test set follow. I agree that it's not always true in practice, and this problem is a bit of a blind spot in ML (at least I'm not aware of any concept/method meant to address this case). | |
| Jan 9, 2022 at 10:38 | comment | added | Amit Keinan | I think it is problematic in real life problems, where the distribution is unknown or may change. In example, corona tests where the sick ratio changes through time. Thanks for the great answer! | |
| Jan 9, 2022 at 10:36 | vote | accept | Amit Keinan | ||
| Jan 9, 2022 at 0:20 | comment | added | Erwan | @AmitKeinan I updated the answer. As I said in the answer I don't know if there is a strong reason one way or the other. But you should probably explain why you think that having performance depend on class distribution is problematic, because from an evaluation perspective I don't think it's so important. | |
| Jan 9, 2022 at 0:16 | history | edited | Erwan | CC BY-SA 4.0 |
added 2371 characters in body
|
| Jan 8, 2022 at 19:15 | comment | added | Amit Keinan | Thanks for the answer, but you didn't address my main point; precision is not an fixed attribute of the model, and therefore it is very problematic. | |
| Jan 8, 2022 at 18:40 | history | answered | Erwan | CC BY-SA 4.0 |