On the consistency of information lters for

A common practice when ltering a case-base is to employ a ltering scheme that decides which cases to delete, as well as how many cases to delete, such that the storage requirements are minimized and the classiication competence is preserved or improved. We introduce an algorithm that rivals the most successful existing algorithm in the average case when ltering 30 classiication problems. Neither algorithm consistently outperforms the other, with each performing well on diierent problems. We also discuss recent work which indicates that ltering can be harmful in some domains, and we conclude that when class deenitions are not made up of homogeneous regions a non-resolvable ambiguity arises between noise and genuine class exceptions. Consistency over many domains, we argue, is very hard to achieve when deploying a ltering algorithm.