Factorial design for selecting an emotion recognition model using artificial intelligence

This article studies the effect of certain supervised training hyperparameters (number of epochs, batch size, and learning rate) on the performance of an artificial intelligence model for recognizing five emotions from facial images. A full factorial design was used, with 60 combinations run 12 times, resulting in 720 experiments developed on Google's Teachable Machine platform. The dependent variable corresponds to the percentage of accuracy in identifying emotions. The data were analyzed using a four-way analysis of variance and post hoc tests. The results show that varying the learning rate from 0.0001 to 0.001 increases accuracy by 46%, using a larger number of batches improves accuracy by 41%, and changing the batch size has the least improvement effect and increases variability. In terms of emotions, neutral was the best identified and sadness the worst. Statistical analysis confirmed significant differences between factor levels and their interactions. It is concluded that properly selecting supervised training hyperparameters is critical to improving the performance of the artificial intelligence model.