Timeline for What are arguments against using the (log-)likelihood as a loss function?
Current License: CC BY-SA 4.0
5 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Nov 23, 2020 at 22:07 | comment | added | Frank Harrell | Don't speak of error functions but rather accuracy scores or loss functions. The fact that many exists doesn't mean that log-likelihood isn't the best, so I'm not clear about your comment. | |
| Nov 23, 2020 at 13:00 | comment | added | Cagdas Ozgenc | @FrankHarrell As a special case, for alphabet size 2 (binary classification), there are many optimal error functions. Not only log-likelihood. | |
| Nov 23, 2020 at 12:36 | history | edited | Sextus Empiricus | CC BY-SA 4.0 |
added 14 characters in body
|
| Nov 23, 2020 at 12:33 | comment | added | Frank Harrell | Well said, and if the dimensionality of the problem increases only slowly as as $n \rightarrow \infty$ maximum likelihood estimates are ultimately optimal. That means that log-likelihood is ultimately an optimal optimality criterion. | |
| Nov 23, 2020 at 12:25 | history | answered | Sextus Empiricus | CC BY-SA 4.0 |