Inproceedings,
Some Probabilistic Modelling Ideas For Boolean Classification In Genetic Programming
Genetic Programming, Proceedings of EuroGP'2000, volume 1802 of LNCS, page 133--148. Edinburgh, Springer-Verlag, (15-16 April 2000)
Abstract
We discuss the problem of boolean classification via
Genetic Programming. When predictors are numeric, the
standard approach proceeds by classifying according to
the sign of the value provided by the evaluated
function. We consider an alternative approach whereby
the magnitude of such a quantity also plays a role in
prediction and evaluation. Specifically, the original,
unconstrained value is transformed into a probability
value which is then used to elicit the classification.
This idea stems from the well-known logistic regression
paradigm and can be seen as an attempt to squeeze all
the information in each individual function. We
investigate the empirical behaviour of these variants
and discuss a third evaluation measure equally based on
probabilistic ideas. To put these ideas in perspective,
we present comparative results obtained by alternative
methods, namely recursive splitting and logistic
regression.
