Enhancing Classification Accuracy Using Stacked Ensemble Learning: A Hybrid Ensemble Strategy

Sultan, Sherif Samir; Ewieda, Mahmoued; Gomaa, Ahmed; Yahia, Mohamed Fathi

doi:10.21608/ajcit.2025.394368.1014

Enhancing Classification Accuracy Using Stacked Ensemble Learning: A Hybrid Ensemble Strategy

Document Type : Research article

Authors

¹ Business Information Systems Department, Faculty of Business Administration, Al RYADA University for science and technology, Cairo, Egypt

² Faculty of Commerce and Administration, University 6 October, Cairo, Egypt

³ Faculty of Computer and Artificial Intelligence, Al RYADA University for science and technology, Cairo, Egypt

10.21608/ajcit.2025.394368.1014

Abstract

Ensemble modeling has become a critical approach in modern machine learning, substantially enhancing predictive accuracy by aggregating the strengths of multiple classifiers while mitigating individual model biases and variance. This study evaluates the effectiveness of a stacking ensemble framework that integrates a Support Vector Machine (SVM) with a Radial Basis Function (RBF) kernel and a Multi-Layer Perceptron (MLP)-based Neural Network (NN). These base models, developed using distinct learning paradigms, exhibit complementary generalization capabilities and are combined into a unified meta-classifier through stacking techniques. The methodology was applied to Fisher’s Iris dataset, a well-established multivariate benchmark widely used in pattern recognition research. The classification pipeline comprised two main phases: the independent development of the base models and the construction of the stacked ensemble. The dataset was partitioned into 80% for training and 20% for testing to evaluate performance consistency. Experimental results indicate that the SVM model achieved a training accuracy of 99.17%, a Matthews Correlation Coefficient (MCC) of 0.9876, and an F1-score of 0.9917. The MLP-based NN attained a training accuracy of 98.33%, an MCC of 0.9754, and an F1-score of 0.9833. Notably, the stacked ensemble model outperformed both base classifiers, achieving perfect test set metrics with 100% accuracy, MCC, and F1-score. These findings confirm the robustness and superior predictive capacity of the stacking ensemble approach over individual models and underscore its potential for constructing high-performing, reliable classification systems.

Keywords

Main Subjects