An Early-Warning Educational System with Small Samples and Across Academic Year Based on Combination of Transformer and XGBoost

Japan’s declining birthrate has been leading to an increasing distribution of academic abilities among university students. To ensure the successful completion of studies, it is crucial to identify students at-risk of academic failure at an early stage and provide them with the necessary support. In conventional teaching systems, this task is highly dependent on teaching experience. Recently developed early-warning systems (EWSs), which are based on the data mining of learning management systems, are built in post-hoc models and lack sufficient generalizability to new academic years. In this study, we present an EWS that combines data augmentation with a transfer learning-enhanced classification model. Through the use of multiple sub-modules, we construct a deep-learning classification model that enhances the generalizability of the EWS. A comparison between a conventional predictive classification model and the presented model shows that our model achieved the optimal overall performance. The stability of this fine-tuned model is verified by the hold-out method. Our EWS is purposefully addressed for difficulties in real-world teaching environments (that is, year-to-year sample domain shifts and small sample sizes); thus, it is robustly adaptable to diverse teaching environments. Teachers can use the recommendations made by the EWS to implement next steps in academic interventions, improve learning strategies, and help students succeed in their studies. All data collection uses non-identifiable information and protects privacy.