An Early Warning Method Combining Autoencoder and Local Attention Mechanism for K-12 Education

As online learning has gained wide attention in K - 12 education, it is more important to identify students with academic risks as early and accurately as possible through early warning models to help teachers carry out personalized interventions and achieve joint efforts between schools and families to improve students' learning performance. Students' online learning behaviors are continuous and time-dependent. The early warning effect can be improved through the analysis of learning behavior patterns. This paper designs an early warning model AE-Local-AT that combines autoencoders and local attention mechanisms. The model eliminates noise in the data and extracts behavioral pattern features through autoencoders, and then captures the behavioral dependencies between adjacent time steps through the sliding window strategy of the local attention mechanism to focus on key features. The test results on K -12 dataset 1 show that AE-Local-AT achieves higher recall, F1-Score, and AUC than the baseline model and other attention mechanisms, and identifies 3/4 of at-risk students 2 units earlier than other models. While having higher prediction indicators, it can reserve more time for teachers and parents to intervene in teaching. This paper further verifies the above results on K -12 dataset 2.