Parameter-Free Imputation for Imbalance Datasets

Class imbalance is a problem that aims to improve the accuracy of a minority class, while imputation is a process to replace missing values. Traditionally, class imbalance and imputation problems are considered independently. In addition, filled-in minority-class values that are substituted by traditional methods are not sufficient for imbalance datasets. In this paper, we provide a new parameter-free imputation to operate on imbalance datasets by estimating a random value between the mean of the missing value attribute and a value in this attribute of the closet record instance from the missing value record. Our proposed algorithm ignores mean of instances to avoid an over-fitting problem. Consequently, experimental results on imbalance datasets reveal that our imputation outperforms other techniques, when class imbalance measures are used.