An oversampling algorithm for high-dimensional imbalanced learning with class overlapping

Most existing standard learning methods suffer from poor performance in high-dimensional imbalanced learning with class overlapping. To tackle this problem, we propose a novel oversampling algorithm that aims to generate a robust ensemble of manifold dimensionality reduction, grid clustering, and information entropy criteria. Instead of simply balancing positive and negative numbers, the algorithm considers the difference in information entropy for interclass, which first reduces the dimensionality by manifold reduction, and then group data utilize grid clustering. Subsequently, calculate the oversampling weight of each group by information entropy and find seed samples based on entropy and neighborhood. Finally, SMOTE based on Beta distribution combined with standard classifiers achieve the rapid and precise classification for high-dimensional imbalanced datasets with class overlapping. Extensive experimental results on 20 real-world imbalanced datasets and compared with eight popular oversampling algorithms show that our proposed algorithm, while achieving good performance in terms of F-measure, G-mean, and AUPRC, can lead to robust performance under high-dimensional and overlapping. It is worth noting that our algorithm substantially reduces the number of synthetic samples against the quantity-balanced oversampling algorithms, and significantly reduces the generation of class overlapping.