Asymmetric misclassification costs and imbalanced group sizes in neural networks for classification

Hundreds of research articles regarding artificial neural networks in solutions to classification problems have been published in the last several years. Although successful applications abound, most researchers fail to explicitly address two important factors that impact the performance of a classification model: imbalanced data and the consequences of unequal misclassification cost. Using both simulated and real datasets, this research explored the effects of asymmetric misclassification costs and imbalanced group proportions on neural network classification performance. Simulated data enable the comparison of neural network estimates of posterior probabilities used in classification to known values for ready analysis. Real data allow comparison with the outcomes of previous studies and with simulated data. The results show that both asymmetric misclassification costs and imbalance group proportions have significant effects on neural network classification performance and that this impact can outweigh the benefit of larger samples. Clearly, when imbalanced data or unequal misclassification costs are ignored, biased results leading to incorrect or meaningless conclusions may be the consequence.