Complexity-Guided Ensemble Learning for Imbalanced Data Classification

Learning from imbalanced multi-class data remains a challenging problem, as standard classifiers tend to be biased toward majority classes. While ensemble methods and resampling strategies have demonstrated strong performance in this setting, most existing approaches do not explicitly leverage data complexity to promote meaningful classifier diversity. This paper introduces the Complexity Guided Ensemble (CGE), an ensemble framework that integrates instance hardness estimation with hybrid resampling. Instance difficulty is quantified using complementary criteria based on error rate, class overlap, and neighborhood disagreement. These measures guide a hybrid over- and undersampling process that generates balanced training subsets emphasizing different regions of the complexity space. By varying a complexity parameter, CGE constructs a diverse ensemble of classifiers trained on subsets with distinct complexity profiles. Experiments on 22 datasets show that CGE achieves competitive or superior performance against 10 state-of-the-art methods, attaining an average macro-F1 score of 0.758 and achieving 167 wins (75.9%), across 220 pairwise Win–Tie–Loss comparisons.