A hybrid multi-criteria meta-learner based classifier for imbalanced data

Abstract

Numerous imbalanced datasets exist in modern machine learning dilemmas. Challenges of generalization and fairness stem from the existence of underrepresented classes with sensitive characteristics. Additionally, defining an optimal criterion is crucial when dealing with imbalanced data. To address this, it is common in practice to employ composite metrics and loss functions that evenly distribute the cost across each class.However, relying solely on a single metric or loss function, frequently yields sub-par outcomes due to the tendency of the optimization criteria to be overly responsive to imbalanced data, resulting in excessive adaptation and flawed generalization. In order to overcome these limitations, various heuristics can be applied to optimize the criterion.To this end, we aim at addressing the problem of imbalanced data classification by leveraging a multicriteria ensemble procedure, the Hybrid Multi-criteria Meta-learner (HML). The suggested approach focuses on optimizing precision, recall together with balanced accuracy in the Multi-Objective optimization phase and brings forth optimal ensembles, from which the concluding Meta-learner can be chosen after a clustering step, based on a multicriteria decision-making aspect contingent on the selected metrics. Comprehensive experiments portrayed the effectiveness of HML, offering both quality assurance and addresses a vast array of imbalanced data problems.