IMOVNO+: A Regional Partitioning and Meta-Heuristic Ensemble Framework for Imbalanced Multi-Class Learning

A research team led by Soufiane Bacha and Huansheng Ning has introduced IMOVNO+, a novel framework designed to solve the persistent and underexplored problem of imbalanced, overlapping, and noisy data in multi-class machine learning. Unlike binary classification, multi-class settings suffer from unclear minority-majority structures and complex inter-class dependencies, which degrade model reliability and generalization. Traditional methods that rely on geometric distances or treat imbalance locally often fail, removing informative samples or generating poor synthetic data. IMOVNO+ proposes a comprehensive, two-level solution that jointly enhances data quality and algorithmic robustness to handle these intertwined challenges effectively.

The framework operates at both data and algorithmic levels. First, it uses conditional probability to quantify sample informativeness, then partitions datasets into core, overlapping, and noisy regions. A novel overlapping-cleaning algorithm combines Z-score metrics with a big-jump gap distance, while a smart oversampling method controls synthetic sample proximity to prevent new overlaps. At the algorithmic level, a meta-heuristic prunes weak classifiers from ensembles to boost overall robustness. Evaluated on 35 datasets (13 multi-class, 22 binary), IMOVNO+ demonstrated consistent superiority, achieving performance gains of 37-57% in G-mean and 25-44% in F1-score for multi-class tasks, and nearing 100% accuracy in binary classification. This breakthrough is particularly impactful for real-world applications like medical diagnosis or fraud detection, where data is often scarce, imbalanced, and noisy due to collection or privacy constraints.