A Weighted Voting Stacking Ensemble for Complex Multi-Class Classification

Traditional stacking ensemble methods have demonstrated strong potential in complex classification tasks. However, conventional implementations often suffer from limited feature utilization and rigid model integration strategies. This paper proposes a novel Weighted Voting Stacking framework that addresses these limitations through three key innovations. First, it introduces a validation-driven dynamic weighting mechanism that adaptively calibrates the contribution of each base model based on its performance. Second, it incorporates a theory-guided hybrid feature generation method, which strategically combines the predictions of base models with a subset of original features selected via mutual information. Unlike simple feature concatenation used in prior stacking approaches, this method ensures maximum complementarity between learned and raw features, thus enhancing the meta-model's discriminative power. Finally, the framework integrates preprocessing techniquesincluding distribution-based scaling and convergence monitoring-to ensure stability and generalization across complex data distributions. Experimental results on a high-dimensional, imbalanced network traffic dataset demonstrate that the proposed method achieves a weighted <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{F 1}$</tex> score of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 5. 9 2 \%}$</tex>, representing a 3.32 % improvement over traditional ensemble methods. These findings highlight the effectiveness of Weighted Voting Stacking in addressing challenges such as feature sparsity, high dimensionality, and class imbalance in multi-class classification tasks.