On the consistency of information filters for lazy learning algorithms

Abstract

A common practice when filtering a case-base is to employ a filtering scheme that decides which cases to delete, as well as how many cases to delete, such that the storage requirements are minimized and the classification competence is preserved or improved. We introduce an algorithm that rivals the most successful existing algorithm in the average case when filtering 30 classification problems. Neither algorithm consistently outperforms the other, with each performing well on different problems. Consistency over many domains, we argue, is very hard to achieve when deploying a filtering algorithm.