Journal

• 목록

Imbalanced data classification is a prevalent challenge in real-world applications. While a conventional sphere-based classification algorithm, random sphere cover (RSC), evenly constructs a set of spheres for two classes in balanced data using a parameter for the minimum sphere size, it struggles with constructing minority spheres in class-imbalanced data. Although RSC can be combined with existing oversampling methods, this approach requires additional hyperparameters, and its effectiveness decreases as the minority size decreases. To overcome these issues, we propose a novel approach that employs the area under the receiver operating characteristic curve (AUC) to construct and expand spheres for minority class. This parameter-free sphere classifier considers both the majority and minority classes simultaneously. We conducted a thorough experiment on both synthetic and 50 real datasets, which revealed that our proposed method outperformed existing various oversampling techniques with the lowest training time.