Online learning environments are dynamic and scattered, making user activity evaluation difficult for e-learning systems. This study proposes an autonomous and intelligent evaluator utilizing Stacking Ensemble Federated Learning (SEFL). First, each compute node trains locally on its dataset using eight CNN architectures (DenseNet169, MobileNetV2, Xception, VGG16, ResNet50, DenseNet121, and ResNetV2). A stacking ensemble combines local models, and the top two create a collaborative local ensemble that interacts with a central node. These ensembles are aggregated by the central node to create a global model, which is repeatedly disseminated until convergence. The SEFL system classifies users' on-screen actions with top-1 accuracy over 94% and an F1-score above 0.92 across test datasets. These findings show that the suggested system can monitor and categorize learner behavior, providing a solid foundation for adaptable, scalable, and intelligent E-learning environments.
