Traditional ensemble techniques such as simple majority and conventional Borda count voting have been widely applied in aggregating predictions from ensemble classifiers. This study introduces a novel preferential voting method to optimize the aggregation process by weighting exponentially performance rank of each deep learning model such as Vision Transformer (ViT). Unlike previous ensemble approaches, it aims to enhance the ensemble performance by tuning the parameter. Using a facial expression dataset consisting of 7 human emotion classes, experimental results demonstrate that the new exponential voting method outperforms traditional voting techniques, suggesting its potential for broader application in ensemble learning for facial classification tasks.