Dynamic F1-score-based voting strategies for multi-class classification: an adaptive ensemble approach for non-linear and imbalanced datasets

Résumé

Classification is a core machine learning task, and ensemble voting methods are widely used to improve predictive accuracy in domains such as medical diagnosis, where class imbalance and non-linear decision boundaries are common. Conventional strategies: Majority Voting (MV), Weighted Voting (WV), and Soft Voting (SV) rely on static or classifier-level weighting schemes that fail to capture per-class differences in classifier reliability. Three dynamic, class-specific voting strategies are introduced: Highest Class F1-Score Voting (HCF1V), Cumulative Class F1-Score Voting (CCF1V), and Enhanced Class F1-Score Voting (ECF1V), each assigning classifier weights based on per-class F1-scores obtained during validation rather than overall performance. The strategies were evaluated through computational simulation on three synthetic non-linear datasets (Gaussian Mixture, Spiral, and Moon) and two real-world medical benchmarks-the Breast Cancer Wisconsin Dataset (BCWD) and the UCI Heart Disease Dataset (UHDD)-using scikit-learn-based classifiers, with statistical significance assessed via Wilcoxon signed-rank tests. ECF1V achieved the highest accuracy across most settings, reaching 98.25% on BCWD and 89.47% on UHDD, outperforming both conventional voting methods and several recently published approaches. These results indicate that class-specific F1-score-based weighting improves ensemble reliability, particularly under class imbalance, supporting its applicability to high-stakes classification tasks such as medical diagnosis.

Citer ce document

Accès au document

Texte intégral en lecture en ligne, réservé aux abonnés SPHAERO et aux membres de l'institution. Se connecter

Auteur(s)

Statistiques

Consultations : 2

Téléchargements : 0