The Good, the Bad and the Incorrectly Classified: Profiling Cases for Case-Base Editing

Case-based approaches to classification, as instance-based learning techniques, have a particular reliance on training examples that other supervised learning techniques do not have. In this paper we present the RDCL case profiling technique that categorises each case in a casebase based on its classification by the case-base, the benefit it has and/or the damage it causes by its inclusion in the case-base. We show how these case profiles can identify the cases that should be removed from a case-base in order to improve generalisation accuracy and we show what aspects of existing noise reduction algorithms contribute to good performance and what do not.