Cluster-based oversampling with area extraction from representative points for class imbalance learning

Class imbalance learning is challenging in various domains where training datasets exhibit disproportionate samples in a specific class. Resampling methods have been used to adjust the class distribution, but they often have limitations for small disjunct minority subsets. This paper introduces AROSS, an adaptive cluster-based oversampling approach that addresses these limitations. AROSS utilizes an optimized agglomerative clustering algorithm with the Cophenetic Correlation Coefficient and the Bayesian Information Criterion to identify representative areas of the minority class. Safe and half-safe areas are obtained using an incremental k-Nearest Neighbor strategy, and oversampling is performed with a truncated hyperspherical Gaussian distribution. Experimental evaluations on 70 binary datasets demonstrate the effectiveness of AROSS in improving class imbalance learning performance, making it a promising solution for mitigating class imbalance challenges, especially for small disjunct minority subsets.