HS-Gen: a hypersphere-constrained generation mechanism to improve synthetic minority oversampling for imbalanced classification

Abstract Mitigating the impact of class-imbalance data on classifiers is a challenging task in machine learning. SMOTE well-known method to tackle this by modifying class distribution and generating synthetic instances. However, most SMOTE-based methods focus phase selection, while few consider generation. This paper proposes hypersphere-constrained generation mechanism (HS-Gen) improve minority oversampling. Unlike linear interpolation commonly used methods, HS-Gen generates instance hypersphere rather than straight line. expands range instances with significant randomness diversity. Furthermore, attached noise prevention strategy that adaptively shrinks determining whether new fall into majority region. can be regarded as an oversampling optimization flexibly embedded methods. We conduct comparative experiments embedding original SMOTE, Borderline-SMOTE, ADASYN, k -means RSMOTE. Experimental results show versions generate higher quality ones. Moreover, these oversampled datasets, conventional (C4.5 Adaboost) obtain performance improvement terms F 1 measure G -mean.