Kernel-based dynamic ensemble approach for classifying imbalanced data with overlapping classes

Abstract In many real-world applications, binary and multi-class classification problems involving imbalanced data and overlapping boundaries present a significant challenge for traditional machine learning algorithms. In this paper we propose a Dynamic Ensemble Selection framework using a Boundary-Aware Kernel (DES-BAK). We investigate the use of ensemble learning approaches to tackle this problem. The aim is to enhance classification tasks based on accuracy, precision, and G-mean by proposing an ensemble of classifiers that leverage different feature representations and classification algorithms. We introduce a novel boundary separation method for the kernel function to separate the imbalanced classes, reduce overlapping, and further improve the ensemble’s performance. The purpose of this method is to divide overlapping boundaries in the classification process. We assess the efficacy of the given method within the framework of imbalanced data in binary and multi-class skewed classification issues with overlapping constraints through Experiments conducted on 15 benchmark datasets. The results demonstrate that the framework surpasses several state-of-the-art methods in terms of classification accuracy. The combination of diverse feature representations, classification algorithms, and the innovative boundary separation approach enhances the ability of the ensemble to handle imbalanced data and overlapping boundaries. These findings showcase the potential of proposed an approach in addressing challenging classification scenarios and contribute to advancing machine learning techniques in real-world applications.