Adaptive Content Recommendation in MOOCs Using Transformer-Based Deep Learning Models

Massive Open Online Courses (MOOCs) increasingly rely on personalized learning technologies to address learner disengagement and high dropout rates. This study proposes a Transformer-based adaptive recommendation model capable of capturing long-range sequential dependencies in learner interaction histories.  Attention-weight analysis reveals pedagogically meaningful behavior, with the model assigning highest importance to Quiz_Attempt (0.241) and Assignment_Submission (0.214) events, while contextual interactions such as video views and page read received moderate weights (0.187 and 0.143). Sequential dependency analysis further shows strong learner behavior patterns, particularly transitions from Video_View → Quiz_Attempt (0.372) and Page_Read → Video_View (0.298). Hyperparameter sensitivity experiments indicate that a configuration of 6 Transformer layers, 12 attention heads, and 256-dimensional embeddings produces the best performance.  Overall, these results demonstrate that Transformer-based models not only improve recommendation accuracy but also enhance explainability through attention mapping and behavioral interpretation. The findings suggest that integrating self-attention mechanisms and behavioral analytics can substantially advance adaptive learning in large-scale online education environments, supporting more personalized and pedagogically aligned learning pathways.