A Closest Fit Approach to Missing Attribute VAlues in Preterm Birth Data

Recently, results on a comparison of seven successful methods of handling missing attribute values were reported. This paper describes experimental results on the three most successful methods out of these seven. Two of these methods, based on a Closet Fit idea (searching in a remaining data set for the closest fit case and replacing a missing attribute value by the corresponding known value from that case) were enhanced by 12 strategies (using two different options for missing attribute replacement, three different interpretations of missing attribute values and two types of rules: certain and possible). Our results show that for a given data set the best method handling missing attribute values should be selected individually, testing the main two methods: Local Closest Fit method with four options: both types of missing attribute replacement and two interpretations of missing attribute values and the Most Common Value for symbolic attributes and Average Value for numerical attributes, both restricted to a concept. All of these methods are local, i.e., restricted to a concept, so it indicates all over again that local approaches are better than global ones.