Hybrid Deep Learning Framework for Student Grade Prediction

This paper constructs a hybrid prediction framework combining Multi-Layer Perceptron, Long Short-Term Memory, and Transformer models to predict students' academic performance. The framework takes the Grades dataset, combines advanced feature engineering technologies such as time attenuation weighting and topic correlation matrix, and uses mean square error (MSE) and R index to evaluate the prediction performance. The results show that the performance of the Long Short-Term Memory (LSTM) model is better than other models, and the minimum mean square error of its test set is 4.4821, proving the LSTM model's effectiveness in capturing students' learning time series patterns. The transformer model also performs well, but the mean square error is slightly higher. In addition, the interpretability analysis of SHapley Additive exPlanations (SHAP) reveals the significant contribution of G2 characteristics to the prediction, which provides a basis for targeted education intervention. This study emphasizes the potential of deep learning in educational data mining and emphasizes the importance of interpretability and feature diversity. Future work may explore the integration of more complex feature engineering technologies and further improve the prediction accuracy by combining the advantages of LSTM and transformer models.