Back Propagation with Balanced MSE Cost Function and Nearest Neighbor Editing for Handling Class Overlap and Class Imbalance

The class imbalance problem has been considered a critical factor for designing and constructing the supervised classifiers. In the case of artificial neural networks, this complexity negatively affects the generalization process on under-represented classes. However, it has also been observed that the decrease in the performance attainable of standard learners is not directly caused by the class imbalance, but is also related with other difficulties, such as overlapping. In this work, a new empirical study for handling class overlap and class imbalance on multi-class problem is described. In order to solve this problem, we propose the joint use of editing techniques and a modified MSE cost function for MLP. This analysis was made on a remote sensing data . The experimental results demonstrate the consistency and validity of the combined strategy here proposed.