Case Representation Issues for Case-Based Reasoning from Ensemble Research

Abstract

Ensembles of classifiers will produce lower errors than the member classifiers if there is diversity in the ensemble. One means of producing this diversity in nearest neighbour classifiers is to base the member classifiers on different feature subsets. In this paper we show four examples where this is the case. This has implications for the practice of feature subset selection (an important issue in CBR and data-mining) because it shows that there is no best feature subset to represent a problem. We show that if diversity is emphasised in the development of the ensemble that the ensemble members appear to be local learners specializing in sub-domains of the problem space. The paper concludes with some proposals on how analysis of ensembles of local learners might provide insight on problem-space decomposition for hierarchical CBR.