ADBoost: boosting with tree-based hybrid-sampling for imbalanced data

Class imbalance remains a pervasive challenge in real-world datasets, frequently biasing predictive models toward majority classes. We propose ADBoost, a unified framework that integrates tree-based hybrid sampling into the boosting cycle. In the oversampling phase, an Adversarial Random Forest (ARF) synthesizes minority-class examples within leaf nodes, preserving local decision boundaries; in the undersampling phase, a Double Random Forest (DRF) removes leaf nodes consisting exclusively of majority-class instances, curbing over-representation. This strategy fosters robust performance under extreme imbalance. During boosting, misclassified examples are reweighted using an adaptive function based on the imbalance ratio, penalizing minority-class errors more heavily. We demonstrate that this weighting is equivalent, at the population level, to a decision threshold adjustment, reducing to standard AdaBoost in the balanced limit. Evaluations on 30 benchmark datasets show that ADBoost enhances minority-class performance and achieves superior overall balance compared to alternative boosting-based methods. An ablation study further confirms that ARF, DRF, and the weighting mechanism each contribute distinctly to performance, with structure-aware sampling playing the larger role and no single component alone sufficient to reproduce the performance of the complete ADBoost framework.