Fig. 2

The partitioning scheme for obtaining balanced training sets, as a part of the hierarchical meta-classification approach. The record set associated with stage 3 is sampled at-random without replacement from the majority class to obtain 7 subsets (shown as white rectangles in the figure). Each subset contains the same number of records as that included in the set combining stages 4 and 5 (grey rectangles). Each of the sampled stage 3 subsets is paired with the set combining stages 4 and 5, thus forming 7 balanced datasets in total, each having a balanced sample of stage 3 and stages 4&5 records