<p>Ensemble learning, which aggregates the predictions of multiple classifiers, has proven to be a robust tool in various supervised classification tasks. Most ensemble learning methods often rely on arithmetically averaging the embedded classifiers’ predicted class probabilities, making these methods vulnerable to biases caused by dominating classifiers or outliers. Meanwhile, methods based on geometric or harmonic means struggle with undefined outcomes due to the possibility of existing zeros among the probabilities predicted by some classifiers. To overcome these drawbacks, a new ensemble approach is proposed based on the Weighted Truncated Harmonic Mean (WTHM). Instead of averaging the entire set of class predicted probabilities for each instance, only those exclusively between 0 and 1 are included in the harmonic mean calculation, and a calibration factor is then applied to proportionally account for any truncation effects. Along with the WTHM, the Coati metaheuristic optimization algorithm is utilized, in order to selectively identify the ensemble’s base classifiers, which optimize the overall classification performance. The classification capability of the proposed approach which will be called Coati-WTHM-Ens is experimentally tested by eighteen benchmarking datasets from UCI machine learning repository, and a survey-based real-life dataset for school dropouts in Egypt. The classification results demonstrate that the proposed Coati-WTHM-Ens model outperforms seventeen state-of-the-art ensemble methods by achieving higher classification performance in the majority of datasets. The model also surpasses the reported UCI baseline metrics and significantly improves classification accuracy for the school dropout prediction in Egypt.</p>