A modified error function for imbalanced dataset classification problem

The objective of learning is to achieve the least error rate. In this paper we proposed a modified cost function as a means to properly measure error rate for imbalanced dataset. Most cost functions apply the same weights to all classes. However, it has been known that for imbalanced problem, the number of instances in the majority class is larger than the minority class. Therefore, the application of equal weight to all classes will significantly lead to improper classification boundary. That is, for most learning model, the minority class would be dominated by majority class which then causes a misclassification on the minority class. The objective of this paper is to find the appropriate parameters to improve MSE cost function based on overlap ratio and class distribution ratio. Back-propagation algorithm with the proposed modified cost function is used to solve two-class classification problem. UCI datasets are used for the experimentation. The results show that the modified MSE cost function provides a better result than the standard one, based on True-positive rate, G-Mean, and F-measurement.