Abstract

Class imbalance, where the minority class contains substantially fewer observations than the majority class, often leads to biased learning outcomes and elevated misclassification rates for minority class instances, making the detection of rare but important cases particularly challenging. To address these issues, this study proposes a novel distribution-aware oversampling approach based on Normalizing Flows. Unlike traditional oversampling techniques such as Random Oversampling (ROS) or the widely used SMOTE method, which may induce overfitting or generate unrealistic synthetic samples, the proposed approach leverages the expressive capacity of Normalizing Flows to accurately learn the underlying probability distribution of the minority class. High-quality synthetic minority samples are then generated from the estimated distribution, capturing both global density characteristics and local structural patterns. This results in improved minority class representation without discarding or distorting majority class information. Extensive simulation experiments across multiple datasets demonstrate that the proposed method outperforms state-of-the-art oversampling techniques in terms of classification accuracy, sensitivity, and robustness.