bits

New essay: Stop worrying about class imbalance

I published a new essay: Stop worrying about class imbalance.

The short version: if you care about calibrated probabilities, log loss already treats the base rate correctly. Resampling can be useful for a deliberate decision objective, but it is not a neutral fix for an imbalanced training set.

Keyboard

gh
go to home
gw
go to writing
gb
go to bits
gp
go to projects
j/k
move focus down/up in list
o
open focused item
t
toggle light/dark theme
?
show this help