A Multi-criteria Evaluation Framework For Facial Expression Recognition Models

Emotions play a pivotal role in the landscape of online learning sessions, and the incorporation of Facial Expression Recognition (FER) algorithms offers valuable insights into students’ learning experiences. However, the majority of current research has mainly focused on evaluating the accuracy and effectiveness of facial expression recognition (FER) models, resulting in an important gap in the field of comprehensive evaluation that prioritizes the crucial aspect of explainable artificial intelligence (XAI). In response, this paper introduces a novel framework tailored for the assessment of FER models within web-based learning environments. A meticulously conducted controlled experiment serves as a robust validation of the proposed evaluation criteria, further reinforcing our approach. Our research results underscore the introduced framework’s capacity to provide a comprehensive, multifaceted assessment of any FER model well-suited for web-based learning environments. The overall robustness and replicability of the experiment are ensured by using publicly available algorithms and real training data, i.e., pre-trained FER models and datasets with labeled facial expressions. However, it is imperative to acknowledge the inherent limitations of our approach, despite these promising findings. Additionally, we outline a pathway for future research endeavors, which includes delving into the domain of global explainability and conducting a nuanced examination of data-based criteria. This exploration notably emphasizes the potential implications of data bias, reflecting the evolving landscape of artificial intelligence.