Multitree genetic programming with spherical-based operators for synthetic minority over-sampling technique in unbalanced data

Unbalanced classification is a critical challenge in machine learning, with broad applications in real-world scenarios. Recent studies have emphasized the potential of Evolutionary Computation (EC)-based approaches, particularly Genetic Programming (GP), as an effective sampling strategy for addressing class imbalance. In contrast to traditional oversampling methods that rely on neighborhood information and predefined structures, GP autonomously selects high-quality instances and evolves structures to generate new ones. However, existing GP-based approaches primarily focus on undersampling, with limited exploration of instance generation. Additionally, the traditional Single-Tree Genetic Programming (STGP) structure struggles to adapt to tasks requiring the generation of multiple candidate datasets. To address these challenges, this paper introduces MTGP-SMOTE, a novel oversampling method based on Multi-Tree Genetic Programming (MTGP). Unlike STGP, which evolves a single tree per individual, MTGP evolves multiple trees within an individual, enabling the generation of diverse new instances while evolving as a complete dataset. The method also incorporates innovative MTGP crossover and mutation operators, designed to enhance exploration by focusing on trees beyond the hemispheres of the target minority class while preserving high-quality individuals throughout the evolutionary process. Experiments on 20 unbalanced datasets demonstrate that MTGP-SMOTE significantly outperforms traditional sampling methods in reducing classifier bias and improving classification accuracy. These results underscore MTGP-SMOTE as a powerful and effective solution for addressing unbalanced classification in machine learning.