Knowledge Intensive Exception Spaces

Abstract

In this paper we extend the concept of exception spaces as defined by Cost and Salzberg (Cost and Salzberg, 1993), in the context of exemplar-based reasoning. Cost et al. defined exception spaces based on the goodness, in terms of performance, of an exemplar. While this is straightforward when using exemplars for classification problems, such a definition does not exist for regression problems. Thus, firstly we define a measure of goodness of an exemplar. We then use this measure of goodness to compare the effectiveness of exception spaces with a variant that we introduce, called Knowledge Intensive Exception Spaces or KINS. KINS remove the restriction on the geometric shape of exception spaces as defined by Cost et al. We provide a rationale for KINS and use a data set from the domain of colorectal cancer to support our hypothesis that KINS are a useful extension to exception spaces.