In online education, recommender systems are essential for guiding students toward courses aligned with their strengths. However, many systems amplify popularity bias, over-recommending dominant domains while neglecting underrepresented ones. This study introduces a neural network-based recommender system to predict student success and deliver equitable course recommendations. Our approach combines a deep learning model leveraging student behavioral data and course attributes to predict success with high accuracy, outperforming standard baselines, and a domain-adaptive re-ranking layer that balances relevance with diversity by boosting recommendations for underrepresented education domains. Ablation studies confirm the critical role of VLE clicks and assessment scores in performance. By embedding fairness through diverse recommendations, this scalable framework shifts from purely accuracy-driven models to inclusive, personalized learning, fostering broader student development in online education platforms. Evaluated on the OULAD dataset, the system achieves an F1-score of 85.91%, AUC of 87.45%, and Precision@5 of 1.0 for sample users, ensuring accurate predictions and relevant recommendations.