Abstract
Although deep learning models, especially the Artificial Neural Network (ANN), are widely used for student performance prediction, their "black-box" nature often leads to unreliable learned relationships that contradict educational domain knowledge, limiting both trustworthiness and further performance improvement. Based on Shapley Additive Explanations (SHAP), this study developed an ANN using a public dataset containing 395 Portuguese high school students' mathematics performance records. The analysis identified key features influencing students' mathematics performance and revealed that the original correlations learned by the ANN were inconsistent with established educational domain knowledge. To address this issue, we proposed the Student Performance Prediction Explanation (SPPE) algorithm for optimizing ANN, which reassessed the contribution of 30 features under the guidance of educational domain knowledge. Both global and local interpretability analyses were conducted to examine the process of importance changes. Furthermore, this study found that after aligning the model with educational domain knowledge, the prediction accuracy of the proposed ANN achieved a 26.9% improvement compared with the original model. In addition, it outperformed some typical traditional machine learning algorithms. Additional experiments further confirmed that the proposed SPPE strategy is applicable to various ANN architectures, supporting its robustness across model structures within this dataset and reinforcing its generalizability and practical value. The findings of this study demonstrated that integrating educational domain knowledge can improve student performance prediction, contributing to the development of interpretable neural network frameworks and offering actionable insights for other educational applications.