Music Learning Path Optimization System Integrating Graph Neural Network

Abstract

Traditional music learning path recommendation systems often fail to model multidimensional correlations within knowledge structures, resulting in incomplete path coverage and limited adaptability. This study proposes a learning path optimization system integrating GraphSAGE and Dueling DQN. A heterogeneous music knowledge graph is first constructed to represent knowledge points, skills, styles, and their prerequisite or coupling relationships. GraphSAGE with an LSTM aggregator is then used to dynamically fuse multimodal features into unified node embeddings. On this basis, the Dueling DQN algorithm uses cognitive state vectors, including mastery level and cognitive load, to optimize path strategies under coverage-gain and load-penalty constraints. Experiments show that the recommended paths achieve 89.6% knowledge coverage with an average length of 12.4 steps for beginners. Compared with standard DQN, Dueling DQN improves coverage by 3.3%. The system increases the path completion rate of beginners to 94.2%, outperforming traditional models such as DySAT, and achieves an average ABRSM score of 128.3. By combining heterogeneous graph modeling with reinforcement learning, the proposed framework improves path coverage, reduces route redundancy, and may provide methodological reference for graph-based optimization in electromagnetic-system training and wireless network planning.