With the development of intelligent education, knowledge tracing and personalized recommendation have become essential techniques to improve learning outcomes. However, existing approaches mainly rely on flat knowledge graphs, which fail to represent the hierarchical relations among chapters, concepts, and sub-concepts. Recommendation strategies often focus only on knowledge coverage, while neglecting learner interests and difficulty progression. This leads to limited prediction and recommendation performance. To address these issues, this study proposed a hierarchical graph attention-based knowledge tracing and personalized recommendation framework (HGAKT + PR). The method constructed a three-level knowledge graph to model hierarchical relationships among chapters, knowledge points, and sub-knowledge points. Multi-head graph attention and a dynamic gating mechanism were applied to learn learners' knowledge states. Shared representation learning was then used to jointly optimize answer performance prediction and personalized recommendation. The recommendation strategy integrated learners' weak knowledge areas, difficulty progression, and interest preferences to improve recommendation relevance and diversity. Experimental results showed that the proposed method achieved stable performance and outperformed representative baseline methods across multiple real-world educational datasets. The model effectively characterized learning states under different learning scenarios and supported personalized learning resource delivery. Furthermore, experiments on multiple datasets and parameter sensitivity analyses consistently demonstrated stable performance under different settings. These findings indicated that hierarchical knowledge modeling combined with prediction-recommendation collaborative optimization enabled more accurate learning state estimation and improved personalized learning resource delivery.
Chen et al. (Fri,) studied this question.