In online learning environments, instructors often lack timely and reliable feedback about students’ engagement states, especially in large-scale or asynchronous courses. Automatic engagement estimation from videos provides a promising way to support instructional interventions. However, in practical deployment, an engagement estimation model should not only predict engagement levels but also indicate whether its predictions are reliable, because ambiguous, noisy, or out-of-distribution videos may otherwise lead to misleading decisions. To address this limitation, we propose an Evidential Gating Mixture-of-Experts (EGMoE) network, which integrates uncertainty quantification into engagement estimation. Specifically, a gating Mixture-of-Experts framework is proposed to effectively capture the spatiotemporal representations in an end-to-end manner, providing an alternative way to extract discriminative spatiotemporal information. Furthermore, the Dirichlet distribution is used to model the probability distribution of the class probabilities, and the Dempster-Shafer Theory of Evidence (DST) is introduced to consider the network prediction as the subjective opinions. A function that collects the evidence leading to the opinions is learned by the gating Mixture-of-Experts framework from videos. The proposed model is evaluated on both DAiSEE and EngageWild datasets, and the experimental results show that the proposed model not only outperforms several state-of-the-art methods, but also can identify the irrelevant testing videos via quantified uncertainty.
