Personalized course recommendation is essential for enhancing learning in Massive Open Online Courses (MOOCs). However, MOOC data are typically sparse and incomplete, limiting the effectiveness of graph-based recommender models. To address this challenge, we propose HetGNN-KGAT, a two-stage framework that integrates Heterogeneous Graph Neural Networks (HetGNN) for imputing missing links and attributes with Knowledge Graph Attention Networks (KGAT) for capturing high-order relations. The contributions of this work are threefold. First, we present a graph-based imputation strategy that mitigates sparsity by enriching MOOC data through HetGNN. Second, we demonstrate that enriching the graph allows KGAT to operate more effectively, highlighting that the novelty lies not in simply combining two models but in strategically enabling KGAT to exploit sparse MOOC data. Third, we establish a systematic knowledge discovery pipeline that explicitly evaluates input data quality (completeness and consistency) alongside downstream recommendation performance (MAP, NDCG, Precision, Recall), providing a comprehensive view of how data enrichment improves recommendations. Extensive experiments on the MOOCCubeX dataset show that HetGNN-KGAT consistently outperforms strong baselines in both simulated and real-world scenarios. The framework achieves up to 23.55% relative improvement in MAP, 11.37% in F1-score, and 27.70% in NDCG under simulated conditions, with corresponding gains of 7.82%, 3.73%, and 9.24% in real-world settings. These results highlight that the proposed framework not only strengthens the accuracy of course ranking but also enhances the overall relevance and reliability of personalized recommendations. These results confirm the value of coupling data-quality enhancement with graph-based recommendation and offer practical insights for advancing adaptive educational technologies.