A Novel Approach for Pruning Training Spaces in Ensemble Classifiers

Ensemble classifiers have gained attention because its high performance in identification and classification problems by combining accurate and diverse classifiers. However, combining all classifiers in an ensemble only add to the computational complexity of the ensemble, leading to the emerging of ensemble pruning paradigm. Numerous state-of-the-art studies explore ensemble pruning techniques, including ranking-based, clustering-based, and optimization-based approaches, which involve pruning either training spaces or trained learners. However, this research found two main issues, including 1) there is no existing studies that considers accuracy and diversity simultaneously to prune training spaces in ensemble, and 2) the potential of using quantum-based optimization approach to improve the ensemble performance by enabling to discover the solution of global optimum, instead of local optimum. Therefore, the contributions of this research are: 1) to propose a pruning approach that considers accuracy and diversity simultaneously to determine the optimal training spaces in ensemble, and 2) to propose quantum annealing as the novel quantum-based optimization method for pruning training spaces in ensemble classifiers. The proposed method was evaluated on 11 datasets from UCI repository. The results showcase the superior accuracy of the proposed method across all datasets, outperforming all benchmark methods. Furthermore, the proposed method exhibits the second-highest average diversity values after AdaBoost. Notably, the pruning process utilizing quantum annealing demonstrates faster computational times compared to particle swarm optimization. These results contribute valuable insights into ensemble classifiers and underscore the potential of leveraging quantum computing techniques for enhanced model performance.