Certain approaches for missing-data imputation propose the use of learning techniques to identify regularities and relations between attributes, which are subsequently used to impute some of the missing data. Prior theoretical results suggest that the soundness and completeness of such learning-based techniques can be improved by applying rules anew on the imputed data, as long as one is careful in choosing which rules to apply at each stage. This work presents an empirical investigation of

more »

... e natural learning-based imputation policies: training rules once and applying them repeatedly; training new rules at each iteration; continuing the training of previous rules at each iteration. We examine how the three policies fare across different settings. In line with the predictions of the theory, we find that an iterative learn-predict approach is preferable.