From radial to rectangular basis functions. A new approach for rule learning from large datasets

Automatic extraction of rules from datasets has gained considerable
interest during the last few years. Several approaches have been
proposed, mainly based on Machine Learning algorithms, the most
prominent example being Quinlan's C4.5.

In this paper we propose a new method to find rules in large
databases, that make use of so-called Rectangular Basis Functions
(or RecBF). Each RecBF directly represents one rule, formulating a
condition on all or a subset of all attributes. Because not all
attributes have to be used in each rule, rules tend to be less
restrictive and result in a more generalizing rule set.

The rule finding mechanism makes use of a slightly modified
constructive algorithm already known from Radial Basis Functions.
This algorithm allows to generate the `network of rules' on-line. It
starts off with large, general rules and specializes them
individually, based on conflicts.

In this paper we present the algorithm to construct the rule base,
discuss its properties using a few data sets and outline some
extensions.