Missing values imputation hypothesis: An experimental evaluation

Abstract

Missing values imputation is a basic strategy to deal with incomplete data. Many developed methods treat filled-in values as if they are original data. The correctness of such hypothesis has not been widely studied. In this paper, a philosophical and experimental study on the hypothesis of missing values imputation is discussed. In the experiments, classification accuracy of three learning algorithms with regard to six incomplete data sets are compared, which indicates that missing values imputation may not always help to improve the learning performance. Learning directly from incomplete data without imputation may reach a satisfying performance. The study not only provides an experimental analysis on missing values imputation, but also presents a new view on rule induction from incomplete data, which is much different from previous standpoint.