Experiments in Guided Class Rebalance Based on Class Structure

Up-sampling and down-sampling are the two most used methods in balancing the data when dealing with two class imbalance problem. However, none of the existing approaches to class rebalance take into account class information (e.g. distribution, within and between class distances, imbalance factor). This study presents initial results of up-sampling methods based on various approaches to aggregation of class information such as spread, imbalance factor and the distance between the classes. Artificially generated data are used for experiments. The performance of each up-sampling method is evaluated with respect to how well the resulting data set reflects the underlying original class distribution, in terms of mean, standard deviation and (empirical) distribution function.