Pruning classification rules with instance reduction methods

Generating classification rules from data often leads to large sets of rules that need to be pruned. A new pre-pruning technique for rule induction is presented which applies instance reduction before rule induction. Training three rule classifiers on datasets that have been reduced earlier with instance reduction methods leads to a statistically significant lower number of generated rules, without adversely affecting the predictive performance. The search strategies used by the three algorithms vary in terms of both type (depth-first or beam search) and direction (general-to-specific or specific-to-general)