A framework to undersample and refine the synthetic minority set

Oversampling the minority class is a popular strategy for coping with the imbalance of datasets. It improves the cognition of the minority points to an admissible extent. Nonetheless, the synthetic minority instances accentuate the overlap between the majority class and the augmented minority class. It is detrimental to the rightful cognition of both classes. To this end, this paper introduces a novel strategy to undersample the synthetic minority set. A multi-armed bandit (MAB) guided protocol is followed to [i] identify the synthetic minority instances that contribute to the increased overlap between the two classes and [ii] subsequently remove (undersample) them iteratively to obtain a refined synthetic minority set. Simulation on synthetic datasets shows that the proposed strategy is successful in increasing the Gromov–Wasserstein distance between the original majority class distribution and the synthetic minority points’ distribution (as compared to the regular oversampled data obtained through state-of-the-art techniques). Empirical evaluation in sixteen real-world datasets, four state-of-the-art minority oversamplers, and two refinement techniques manifest the competence of the proposed strategy over baseline results and against the two competing methods. The proposed strategy has improved the performance of the majority class without bringing down the minority class’s performance and can be incorporated in sensitive real-world domains. • This work introduces a novel technique to undersample the synthetic minority set. • This work reduces the overlap between the synthetic minority set and the majority class. • The principles of Multi-Armed Bandit are used to evaluate the overlap caused by synthetic minority instances. • Empirical results show that incoporation of proposed protocol reduces the class overlap.