Abstract
The COVID-19 pandemic necessitated a shift towards remote learning, posing challenges for educators in assessing student impact. While deep learning offers solutions for face expression recognition (FER) in online education, existing evalua- tion methods often overlook the underlying decision-making processes. This research addresses this gap by proposing a robust, theory-driven, and case-oriented framework for preliminary FER model selection. It prioritizes accuracy, validity, and trustworthiness beyond traditional metrics by uncovering inherent biases and assessing algorithmic interpretability and generalization. A controlled experiment validates these criteria within the Design Science Research framework, facilitating further refinement of the evaluation tool. This novel approach prioritizes reliability and interpretability over mere predic- tive accuracy, filling a critical gap in existing FER systems. By integrating conformal prediction techniques, the framework enhances uncertainty quantification, allowing educators to understand the confidence levels associated with FER outputs. By advocating for transparency and clarity, it aligns with the evolving legal and societal emphasis on responsible AI development. This framework enables FER models to provide unbiased insights into students’ online learning experiences, encouraging ethical considerations in FER research. The findings serve as a valuable resource for researchers and developers, fostering trust and comprehension among educators in web-based learning environments.