Abstract: To address issues of sparse interaction data and suboptimal learning resource planning in learning path recommendation, a reinforcement learning (RL)-based model named MBRL4LP was proposed, incorporating multi-behavior user data. First, user behavior data was categorized and embedded as entity nodes into a course knowledge graph. An attention-based Graph Convolutional Network (GCN) was used to capture multi-source heterogeneous features. Second, three data augmentation strategies were designed from both behavioral and learning resource perspectives. Contrastive learning was applied to learn representations from augmented data, which were integrated into RL as supplementary information. Finally, individual differences were fully considered. A personalized learning path recommendation framework was built using a deep Q-network, and a dual reward function—considering both knowledge concepts and path structure—was designed to guide model convergence. Experiments compared MBRL4LP with existing models (e.g., LPG, KTKDM, LSTMPR) on four real-world courses from MOOPer and MOOCCubeX datasets. Results show that MBRL4LP achieves at least 7.42%, 5.97% and 6.15% improvements in precision, recall, and F1-score on MOOPer, and at least 4.75%, 6.62% and 6.42% on MOOCCubeX, respectively. Parameter sensitivity analysis and ablation studies validate model effectiveness of MBRL4LP.

Key words: Graph Convolutional Network (, GCN)