Pruning and Exclusion Criteria for Unordered Incremental Reduced Error Pruning

摘要: Incremental reduced error pruning is a technique that has been extensively used for efficient induction of ordered rule sets (decision lists). Several criteria have developed regarding how to prune rules and whether or not exclude generated rules. A version incremental unordered presented, the appropriateness previously proposed novel investigated. It shown when inducing sets, where Bayesian framework combine predictions from multiple rules, could lead exclusion possibly beneficial as well inclusion harmful Two alternative are introduced, one based on likelihood ratio margin. An empirical evaluation 34 datasets shows significantly outperform employed using margin-based being slightly ahead criterion.