Dynamic integration of preference and knowledge status for knowledge concept recommendation

Knowledge concept recommendation (KCR) has recently emerged as a crucial task for achieving personalized learning within online education systems. Existing studies basically construct heterogeneous information networks (HINs) from students’ learning behaviors, and leverage graph representation learning to model their interests, i.e., preference status. However, current approaches overlook two critical aspects. First, students’ learning motivation is inherently dynamic, yet existing methods typically operate on static HINs, failing to capture the temporal evolution of learning behaviors. Second, learning motivation stems not only from personal interests but also from knowledge deficits, which can be inferred from learners’ knowledge status through their assessment behaviors. Consequently, accurately modeling students’ learning motivation requires dynamically capturing both their preference status and knowledge status. To address these limitations, we propose DISKRec, a novel model that dynamically disentangles and integrates students’ preference and knowledge status for knowledge concept recommendation. Specifically, we first construct a continuous-time heterogeneous information network (CTHIN) based on students’ hybrid behaviors, naturally preserving the underlying temporal dynamics and complex behavioral patterns. Then, we develop a dual dynamic graph neural networks (Dual-DGNNs) module to disentangle preference status and knowledge status from the CTHIN. To dynamically integrate these status, we specially design both implicit dual-state integration and explicit status integration mechanisms. Additionally, we adopt a live-update strategy to improve the training efficiency of DISKRec, substantially reducing GPU memory consumption without sacrificing performance. Extensive experiments demonstrate that DISKRec significantly outperforms existing state-of-the-art KCR models.