I have a binary classifier whose accuracy is $a$, tested in equally distribute samples, so $p(0)=p(1)=0.5$.
If $p(0)$ or $p(1)$ changes to some arbitrary values, is there a way to quantify the performance of my classifier?
Asked
Active
Viewed 10 times
0
-
There are so many ways. Might I recommend starting by reading the Wikipedia article on evaluation of binary classifers then refining the question after that based on what's not clear? – Arya McCarthy Dec 14 '23 at 06:48
-
1@AryaMcCarthy: I would very much counsel against following that Wikipedia article. It considers "hard" 0-1 classification, which may not be what OP is asking for, and where all recommendations suffer from the same issues as accuracy does. – Stephan Kolassa Dec 14 '23 at 07:52
-
1Accuracy is dangerous and misleading, especially - but not only - for "unbalanced data". It is much better to use probabilistic classifications and assess them using proper scoring rules. More info at the tag wiki. – Stephan Kolassa Dec 14 '23 at 07:55