Cost-Sensitive Imputing Missing Values with Ordering

Missing value is an unavoidable problem when dealing with real world data sources, and various approaches for dealing with missing data have been developed. In fact, it is very important to consider the imputation ordering (ordering means which missing value should be imputed at first with the help of a specific criterion) during the imputation process, because not all attributes have the same impact on the imputation results. Usually, the higher correlation between the non-target attributes and the target attributes, the more important the attribute is. On the other hand, imputation ordering is important for reducing costs when we impute a missing value involving costs, including imputation costs and other costs. However, to our knowledge, there are no methods of imputation ordering dedicatedly proposed for missing data imputation, so as to enhance the performance and minimize the imputation cost. There are only few reports on improving the classification accuracy by ordering, for example, Claudio (2003), Numao (1999), and Estevam (2006). In this paper we present two strategies with imputation ordering to minimize imputation cost and improve the accuracy. One is called incremental iterative method, in which each last imputed information are added to training set for imputing the remained missing values, and it is repeated until the accuracy doesn't increase again. The other is the iterative method, in which each missing value is imputed with all information of the dataset including the instances with