Enhanced Modelling Performance with Boosting Ensemble Meta-Learning and Optuna Optimization

Abstract Improving modeling performance on imbalanced multi-class classification problems has continued to attract attention from researchers considering the critical and significant role such models should play in mitigating the prevalent problem. Ensemble Learning (EL) techniques are among the key methods utilized by researchers as they are known for robust and optimal performance when implemented for classification tasks. This study implements optimized state-of-the-art EL algorithms for imbalanced multi-class classification using engineered data. The optimized models are utilized to construct a Hybrid Pre-Stack Ensemble ( Hp-SE ) baseline model and 3-Model Stack Ensemble Meta-Learning Architecture (3-MoSELA) implemented to enhance classification performance on the engineered data. The models are assessed using computational statistics and loss performance. Results indicate improved performance by the proposed 3-MoSELA technique with MoSELA models achieving up to 26.7x faster training time beyond the optimized models, reduced loss by more than 80% for all MoSELA models and up to 89.66% by the best-performing model, indicating the method’s potential as a viable method which can be implemented for diverse multi-class meta-classification problems and future related research considerations.