A Stacked Ensemble Deep Learning Recommender System for Mitigating Data Sparsity Challenges in MOOCs

Abstract

MOOCs have substantially increased educational accessibility; however, the abundance of available courses frequently results in information overload, which is a major factor in learner attrition. Although recommendation systems are critical for personalizing learning pathways, conventional methods often underperform in the presence of severe data sparsity and cold-start scenarios. To address these challenges, this study introduces a stacked ensemble DNN architecture that combines CNN, BiGRU with an attention mechanism, and Transformer encoders. Utilizing pretrained embeddings, the proposed ensemble framework captures both localized textual features and long range semantic dependencies from user reviews, enabling the construction of detailed user profiles even with limited historical interaction data. When evaluated on a comprehensive MOOC dataset, the model demonstrates superior performance compared to established baseline methods including MF, NCF, KNN, and CF achieving near-optimal results with Hit Ratios of 0.98840.9900 and nDCG of 0.9997–0.9998 for top-K recommendations. Although this offline evaluation indicates high recommendation accuracy, which is crucial for alleviating information overload and reducing student attrition, longitudinal deployment is necessary to directly assess real-world course completion rates.