A closest fit approach to missing attribute values in preterm birth data

Abstract

In real-life data, in general, many attribute values are missing. Therefore, rule induction requires preprocessing, where missing attribute values are replaced by appropriate values. The rule induction method used in our research is based on rough set theory. In this paper we present our results on a new approach to missing attribute values called a closest fit. The main idea of the closest fit is based on searching through the set of all cases, considered as vectors of attribute values, for a case that is the most similar to the given case with missing attribute values. There are two possible ways to look for the closest case: We may restrict our attention to the given concept or to the set of all cases. These methods are compared with a special case of the closest fit principle: Replacing missing attribute values by the most common value from the concept. All algorithms were implemented in system OOMIS. Our experiments were performed on preterm birth data sets collected at the Duke University Medical Center.