# ICS 269, Spring 1998: Theory Seminar

## 15 May 1998:

Machine Learning of Classifications via Generalized Linear
Models

Kent Martin, ICS, UC Irvine

We describe a general-purpose classification learning algorithm
based on Generalized Linear Models, and evaluate that algorithm
(GLOREAL, the Generalized Logistic Regression Algorithmic Learner)
as a competitor to ordinary decision tree learners such as C4.5.
GLOREAL is based on a generalization of logistic regression and it
also uses some non-conventional methods for model selection that
are adapted from other work at the interface of machine learning
and statistics. Empirical results show that GLOREAL is 4-5% more
accurate than C4.5 on the average. GLOREAL's success is due in part
to learning and aggregating multiple classifiers for a problem,
which is more effective in GLOREAL than it is in bagging decision
trees. Another important factor in GLOREAL's success is that it has
a richer and more effective representation language than ordinary
decision tree learners, particularly with regard to numeric
attributes.