On the Process of Making Descriptive Rules

Abstract

The automatic inductive learning of production rules in a classification environment is a difficult process which requires several considerations and techniques to be studied. This is more noticeable when the learning process is applied to real world domains. Our goal is to focus and study some of the most important problems related to the automatic learning of production rules as well as to provide some tools for dealing with these problems. We first consider the data representation problem. Four different types of data are proposed. We then deal with the unsupervised case in which the data are observations of objects in the world and we pose three alternative mechanisms for clustering. If the data set contains examples and counter examples of some world concepts, the learning is called supervised. Within supervised learning we find the data redundancy problem. Two sorts of redundancy are studied: the one which is concerned with the set of examples, and the one which is concerned with the set of example descriptors.

Before we generate rules that describe the domain which is represented by the input data, we analyze the set of conditions which will be the basis of our rules. These conditions are called selectors and they enable us to control more directly the semantics of the induced rules. We have implemented several algorithms that generate selectors automatically and we have tested them together with four new rule generation algorithms. The results obtained are compared with those other results produced by other classical rule learning methods such as cn2 and c4.5rules.