A new reparation method for incomplete data in the context of supervised learning

Real-world data is often incomplete. There exist many statistical methods to deal with missing items. However, they assume data distributions which are difficult to justify in the context of supervised learning. In this paper we propose a new method of repairing incomplete data. This technique is a variation of a general strategy, here called local imputation. It repairs incomplete records, only when this is reasonable. It is able to identify wrong tuples. It is more general than other similar methods, because of a parametric similarity function. Finally, it also works with noisy data sets.