An LSTM-TMSGCN-MHSA Model for Student-Course Level Prediction of Academic Risk: Integrating Temporal Dynamics and Inter-Course Dependencies

Accurately identifying the risk of academic failure at the student-course level is fundamental to data-driven early warning and precision student management, which requires a subsequent decision-layer aggregation to inform student-level interventions. However, existing methods fail to jointly capture temporal grade evolution, course prerequisite dependencies, and cross-cohort distribution shifts, limiting their predictive efficacy. To address this, we propose a hybrid model integrating Long Short-Term Memory (LSTM) with a Transferable Multi-Scale Graph Convolutional Network (TMSGCN). Experiments on real records from 17 cohorts across five engineering majors at a Chinese university show that our model achieves strong predictive performance in internal cohort evaluations. In a temporal hold-out test on the 2023 cohort, it attains an AUC of 0.946 and a false-negative rate of 6.7%, highlighting cross-year generalizability. The model consistently outperforms the re-implemented baseline in both predictive accuracy and stability. Interpretability analyses further validate the rationality of its decision-making. The framework shows promising internal transferability within the studied institution.