Predicting and Interpreting Student Academic Performance: A Deep Learning and Shapley Additive Explanations Approach

Predicting students' performance in high-risk exams, such as the baccalaureate, is essential for early identification of at-risk students and designing targeted interventions. This study introduces a deep learning approach to predict final baccalaureate outcomes among Moroccan high school students based on their current performance in the first semester and previous academic achievements. The dataset comprises 182.636 records containing demographic, socioeconomic, and prior academic performance features. We used a neural network model to predict the cumulative grade point average (CGPA). In the testing set, the model achieved a Mean Squared Error (MSE) of 0.258 and a Mean Absolute Error (MAE) of 0.392. Moreover, the model explains 72.3% (R2 score = 0.723) of the variance in the target variable (CGPA), capturing a significant portion of the underlying relationships in this dataset. We also integrate the SHapley Additive exPlanations (SHAP) tool to enhance model interpretability. The SHAP analysis highlights that academic performance, particularly on the regional exam and first-semester overall average, is the most important factor in predicting students' CGPA; poverty and class size also play a role. This work emphasizes the potential of combining deep learning models with interpretability tools to provide actionable insights in educational settings.