Design and Optimization of a Personalized Resource Recommendation Engine Integrating Graph Neural Networks and Course Knowledge Graphs

Abstract

To address inaccurate user-interest modeling, insufficient knowledge-association mining, and cold-start problems in online educational resource recommendation systems, this paper proposes a personalized resource recommendation engine integrating graph neural networks with course knowledge graphs. The method is suitable not only for general online education but also for specialized engineering domains, including electromagnetic waves, antennas, propagation, and smart electronics, where prerequisite relations and resource dependencies are highly structured. First, a heterogeneous graph model of user-resource-knowledge nodes is constructed by fusing structured information from course knowledge graphs with user-interaction data. Second, an attention-based graph convolutional network aggregates user-interest and knowledge-association features through multi-head attention. Finally, contrastive learning strategies are introduced to optimize model training and reduce the influence of data sparsity. Experiments on the public Coursera dataset and a self-built university course dataset show that the proposed engine outperforms traditional collaborative filtering, GCN, and KGCN models. Precision@10 improves by 8.3%-15.6%, Recall@10 improves by 9.1%-17.2%, and NDCG@10 improves by 7.8%-14.9%. The results demonstrate that combining graph-based knowledge representation with attention and contrastive learning effectively improves recommendation precision and personalization, offering a practical route for electromagnetic engineering learning-resource organization.