A Cost-Sensitive Approach applied on Shallow and Deep Neural Networks for Classification of Imbalanced Data
International journal of intelligent engineering and systems · 2023
In this study, we propose a cost-sensitive learning approach applied on neural networks to deal with classification under imbalanced domains.Our approach is able to automatically learn robust features for both frequent and rare classes by automatically assigning misclassification penalties to each class based the frequency of occurrence of that class.This approach is investigated in the context of shallow networks (multi-layer perceptrons) and deep networks (convolutional neural networks).Moreover, it offers not only a better convergence but also a faster convergence since it can boost optimization by increasing weight gradients which are getting small due to their fitting to the frequent classes.Extensive experiments were carried out on one-and two-dimensional datasets.Running experiments using several loss functions showed the efficacity of our approach on loss functions which do not have probability estimates.Additionally, our approach achieved a good performance compared to common undersampling and oversampling methods as well as models based on generative adversarial networks.