Learning algorithms aim for accuracy of classification but this depends on a choice of heuristic metric to measure performance and also on the proper consideration and addressing of the important requirements of the classification task. This paper introduces a framework, MV Gen, to implement different training heuristics capable of inducing the training algorithm that can provide the desired results while negating detrimental aspects of a training set imbalance. Our experiments indicate that successful classifiers can indeed be built to specialize on the minority class within an imbalanced data set.