The effect of pruning on the efficiency and effectiveness of hybrid imbalanced multiclass classification models

Hybrid models are recognized as one of the most effective approaches to address the imbalanced data problem. In these models, data-level methods such as over-sampling are combined with algorithm-level methods, such as ensemble approaches. However, the resulting models can face challenges concerning inefficiency and ineffectiveness. A solution to tackle these issues is proposed in this paper, which includes a novel weighted F1-ordered pruning technique integrated with two state-of-the-art hybrid models, Balanced Bagging and Balanced One-versus-One. Unlike prior hybrid models designed primarily to address the binary imbalance problem, the proposed approach is specifically designed to tackle the challenging multi-class classification imbalance problem. An extensive experimental evaluation and statistical validation were conducted, and demonstrated that the Pruned Balanced Bagging ensemble remarkably outperforms the considered hybrid models.