Adaptive Ensemble Weighting with Local Competence for Multi-Class Imbalanced Classification
Abstract
Multi-class imbalanced classification remains difficult because minority classes can be poorly recognised even when aggregate performance appears acceptable. Many imbalance-handling methods still rely on a fixed technique, a single selected technique, or one level of adaptation, despite the fact that technique suitability changes with both dataset characteristics and local decision regions. This article presents Hybrid Adaptive Ensemble Weighting with Dynamic Ensemble Selection for multi-class imbalanced classification. The model combines two sources of evidence: dataset-level suitability estimated from meta-features, and instance-level competence estimated around each test sample. Adaptive Ensemble Weighting uses kernel similarity over 13 meta-features to estimate the suitability of 10 imbalance-handling technique pipelines, while Dynamic Ensemble Selection measures local classifier competence from validation-neighbourhood behaviour. Both weight vectors are then combined during prediction. Evaluation was carried out on 30 OpenML benchmark datasets with 3 to 20 classes, 160 to 10,000 instances, 4 to 1,300 features, and imbalance ratios from 1.9:1 to 567.2:1. Under Leave-One-Dataset-Out validation, using macro F1-score as the primary metric, Hybrid AEW-DES achieved the highest average macro F1-score of 0.7308 and the best average rank of 3.83 among the 10 evaluated methods. AEW ranked second with an average macro F1-score of 0.7245 and exceeded the Oracle single-technique reference score of 0.7188. The Friedman test confirmed significant differences among methods (p = 0.0020). Wilcoxon signed-rank tests showed significant AEW improvements over Baseline, SingleSelect, KNORA-E, and Stacking, while Hybrid AEW-DES significantly outperformed DES-LA and KNORA-E. These results indicate that combining dataset-level meta-feature weighting with instance-level local competence can improve adaptive technique combination within the evaluated benchmark setting.