A General Method of Fuzzy Instance-Based Learning
Abstract
A method of representing a data set by a smaller set of fuzzy rules is described. The rules are generated from labeled examples that consist of nominal-valued and/or real-valued attributes. By pruning attributes directly from the examples, the size of the data set is reduced through the use of data-driven constructive induction. The system produces two different forms of rules, those that may be analyzed by a human and those that may be incorporated directly into an expert system. Actual rules and rule performance are discussed with respect to the Iris and Promoter domains. The performance of the rules in these domains is comparable to performance reported by other systems in the literature. The rules themselves and/or the methods used to produce them are arguably better than the rules and methods of other systems.