A Deep Learning Regression Model for Predicting Students’ Performance with Shapley Additive Explanations-based Interpretability

Predicting student performance is crucial for identifying at-risk students, guiding educational interventions, and improving student outcomes. In this study, we developed a deep learning regression model to predict students’ performance in Moroccan baccalaureate exams and evaluated the model’s reliability using SHAP (SHapley Additive exPlanations). Our model was trained and evaluated with a large dataset of 342.268 students. The model achieved notable performance, with a Mean Squared Error (MSE) of 0.238, a Root Mean Squared Error (RMSE) of 0.482, and a Mean Absolute Error (MAE) of 0.377 on the test set. An R-squared score of 0.741 indicates that the model explains 74.1% of the data’s variance on the test set, highlighting its reliability in real-world educational contexts. Additionally, the Mean Bias Error (MBE) was near zero (0.0027 on the test set), demonstrating that the model’s predictions are unbiased and neither systematically overestimate nor underestimate student performance. We assessed the importance of features using permutation-based analysis and SHAP values to understand which features most significantly influenced predictions. Results highlighted that the "Regional Exam" score was the dominant factor, followed by features such as "first-semester average" and "Subject-related continuous assessment scores." This study demonstrates the potential of deep learning models in educational prediction and the importance of explainable AI(XAI) in interpreting complex model predictions and supporting educators and policymakers in decision-making.