A model usually has one threshold, for example 0.5, if anything is greater than 0.5 we predict it as 1 vice versa.
However, there are some features that impacts the output probability. Let's say feature a
can equal to either 0
or 1
. If we split the output results between 0
and 1
, the thresholds gathered for 0
greatly differs from 1
, meaning that the thresholds can be much greater than the probabilities that occurs in 0
. Hence that anything predicted with 0
is automatically equal to 0
class, but in reality that is not true.
Is it logical to create two separate thresholds for the 0
and 1
input variable or maybe just create two individual models instead?