This study presents three adaptive, class-specific voting strategies for ensemble classification in settings with non-linear decision boundaries and class imbalance. The proposed methods—Highest Class F1-Score Voting (HCF1V), Cumulative Class F1-Score Voting (CCF1V), and Enhanced Class F1-Score Voting (ECF1V)—use per-class F1-scores computed on validation data to determine classifier weights for each target class. The core idea is to replace static or classifier-level weighting with a fine-grained measure of per-class reliability, with the aim of improving ensemble predictions in challenging classification tasks such as medical diagnosis.
Common ensemble voting schemes include Majority Voting (MV), Weighted Voting (WV), and Soft Voting (SV). These methods typically apply static rules or classifier-level weights that reflect overall performance but do not account for per-class variation in classifier strengths and weaknesses. In domains where class prevalence is unbalanced and decision boundaries are non-linear—conditions frequently encountered in medical datasets—such global weighting can produce suboptimal ensemble decisions. The authors argue that a class-specific weighting scheme based on the F1-score can better capture important differences in classifier reliability across classes.
Three dynamic voting strategies are introduced:
Highest Class F1-Score Voting (HCF1V): assigns voting weight for each classifier based on the classifier’s highest F1-score across classes observed in validation.
Cumulative Class F1-Score Voting (CCF1V): aggregates class-level F1-scores to form cumulative class-specific weights that reflect broader per-class performance across classifiers.
Enhanced Class F1-Score Voting (ECF1V): a refined weighting approach that further adjusts per-class weights using validation-derived F1-scores to emphasize classifiers that are consistently strong for specific classes.
Each method relies on F1-scores computed per class during validation rather than a single overall F1 or accuracy metric. This enables the ensemble to preferentially rely on classifiers that demonstrate higher precision and recall for particular classes, which is especially relevant for imbalanced class distributions.
The strategies were evaluated using computational simulation across multiple datasets chosen to represent non-linear and imbalanced conditions. Three synthetic datasets with non-linear structures were used: Gaussian Mixture, Spiral, and Moon. Two real-world medical benchmark datasets were included: the Breast Cancer Wisconsin Dataset (BCWD) and the UCI Heart Disease Dataset (UHDD). These medical datasets were selected as representative high-stakes diagnostic problems where improved ensemble reliability is clinically meaningful.
Experiments were conducted using scikit-learn-based classifiers. The study compared the three proposed dynamic voting methods against conventional voting approaches (MV, WV, SV) and several recently published methods. Statistical significance of comparative results was assessed using the Wilcoxon signed-rank test. The abstract reports overall experimental outcomes; specific classifier types, hyperparameter settings, cross-validation folds, and full statistical tables were not reported in the abstract and require the full text for complete reproducibility.
Across most experimental settings reported, ECF1V achieved the highest accuracy. On the two medical benchmarks, the abstract reports ECF1V accuracies of 98.25% on BCWD and 89.47% on UHDD, outperforming conventional voting schemes and several recent methods referenced by the authors. These results support the conclusion that class-specific, F1-score-based weighting enhances ensemble performance in non-linear, imbalanced classification tasks.
The authors conclude that dynamic, class-specific weighting based on validation F1-scores increases ensemble reliability and is particularly beneficial under class imbalance. Given the improved accuracies reported on breast cancer and heart disease benchmarks, the approach is presented as applicable to high-stakes clinical classification tasks where per-class errors have different clinical consequences. The focus on F1-score weighting aligns with the need to balance precision and recall in diagnosis-oriented models.
The authors declare no competing interests. The abstract summarizes the methods, datasets, and headline results, but does not include full implementation details such as the exact list of base classifiers, hyperparameter settings, or exhaustive statistical output. Those reproducibility elements are not reported in the source abstract and would require consultation of the full article for complete replication.