Stack generalization-based hybrid ensemble classifier for imbalanced data

Suyash Kumar

doi:10.1007/s10115-026-02752-y

Return

Stack generalization-based hybrid ensemble classifier for imbalanced data

2026-04-06

PRE

Suyash Kumar *

Prabhjot Kaur

Anjana Gosain

DOI：10.1007/s10115-026-02752-y

Original

Original request for help

Save

Abstract

En 中文

An imbalance dataset where one class has significantly more instances than another often causes traditional classifiers to struggle in predicting the minority class correctly. This paper introduces a new ensemble method called Random Balance Stacking (RBStack) to address this issue in binary classification. RBStack uses a two-level stacking approach where we use five classifiers (Adaboost, K nearest neighbor (KNN), Random Forest (RF), Bagging, and Extreme Gradient Boost (XGB)) at base level and RF as the meta-classifier. The process starts by generating a random number that determines the new size of the minority class, ensuring it falls between the original size of minority and majority class. RBStack uses SMOTE-SVM (Synthetic Minority Oversampling Technique with Support Vector Machine) to generate synthetic minority instances. To balance the dataset, it also randomly removes instances from the majority class, maintaining the dataset's original size. Experiment was conducted on 47 imbalanced datasets from the KEEL, UCI, and NMDP repositories, with imbalance ratios ranging from 1.4 to 129.4. We compare the proposed approach with 11 other classifiers using five performance metrics, AUC (Area under the curve), f1-score, g-mean, kappa, and Matthews Correlation Coefficient (MCC). The results show that the RBStack outperforms all other classifiers. The results are also validated by the statistical test Wilcoxon Signed-Rank Test.

Keywords:

Bagging

Boosting

Class imbalance

Ensemble

Oversampling

Stacking