Multi-level Enhanced Contrastive Learning for the Prediction of Learning Effectiveness in MOOCs

With the development of Massive Open Online Courses (MOOCs) platforms, accurately predicting student learning outcomes has become particularly important. To address the limitation of representation learning caused by sparse supervision signals, contrastive learning is introduced as a powerful method, utilizing self-supervised signals between training samples to learn better representations. However, existing self-supervised contrastive learning methods heavily rely on historical performance and behavioral data, overlooking semantic information, cognitive states, and emotional responses. To address the above limitations, we propose a Multi-Level Enhanced Contrastive Learning framework (MECL), which aims to redefine the self-supervised learning process from the perspectives of label perception and structural reconstruction. and learn better representations by exploiting self-supervised signals between training samples. Specifically, we treat the parameters of the classifier as augmented samples associated with different labels, injecting label information into the feature space to obtain enhanced views. In addition, to balance the robustness and generalization ability of learning representations, we designed a structural enhancement mechanism to guide the model in reconstructing multiple feature representations. The experimental results on a real-world MOOC platform dataset demonstrate that the model outperforms the best existing methods in terms of prediction accuracy and generalization across different courses, providing new insights for educational interventions and personalized learning path design.