Ensemble convolutional neural network modeling based on a cost-sensitive approach

Addressing imbalanced data is one of the most significant challenges in machine learning. This research introduces the Cost-sensitive AdaBoost-CNN, a novel machine-learning approach that combines the strengths of cost-sensitive techniques, adaptive boosting, and convolutional neural networks. This method mitigates the misclassification of underrepresented categories in imbalanced data classification. Building on traditional AdaBoost, our approach integrates the principles of cost-sensitive normalization into the AdaBoost algorithm. This study further validated the approach using Fashion-Mnist and CIFAR-10 datasets. The proposed method consistently outperformed the standard AdaBoost-CNN, Integration Tree, AdaBoost-SVM, and other methods, especially for underrepresented class samples. A comparative experiment involving AdaBoost-CNN and classical neural networks, such as Lenet-5 and Alexnet, was also conducted on the PneumoniaMnist medical image dataset. The proposed method exhibited superior accuracy in classifying minority sample data such as specific patient categories. Additionally, on the multi-category retinal disease OCTMnist dataset, the approach demonstrated significant improvements in Recall, F1-Score, G-mean, and P-mean metrics, especially for the 'drusen' sample data when compared to AdaBoost-CNN.