Counterfactual-based minority oversampling for imbalanced classification

A key challenge of oversampling in imbalanced classification is that the generation new minority samples often neglects usage majority classes, resulting most sampling spreading whole space. In view this, we present a framework based on counterfactual theory. Our introduces objective by leveraging rich inherent information classes and explicitly perturbing to generate territory It can be analytically shown satisfy minimum inversion. Therefore, them are located near decision boundary. The empirical evaluation six benchmark datasets shows our approach clearly outperforms state-of-the-art methods.