Iterative rule induction methods

摘要

We examine heuristic techniques for inducing production rules to cover artificially generated boolean expressions with irrelevant noise attributes. The results of different rule induction methods are compared, and it is shown that an iterative tree-based single-best-rule technique performs best on a set of widely-studied applications. We also introduce a new class of iterative Swap-1 rule induction techniques that also solve these problems. While the primary focus is on rule-based solutions, the results for k-nearest neighbor methods and back-propagation neural networks are also reviewed. The results cannot be immediately extrapolated to the more general class of problems with unknown distributions and numerical variables. However, they do offer some comparisons of the effectiveness of competitive rule induction methods that use substantially the same representation.