Enhancing Student Performance Prediction with ANN-Based Transfer Learning

Predicting student performance in higher education is challenging when data distributions differ across cohorts or programs. This paper proposes an adaptive transfer learning framework to improve prediction accuracy on a student dataset with simulated domain shifts. The dataset contains demographic, academic, and macroeconomic features for university students, with the target outcome indicating whether a student graduated, dropped out, or is still enrolled. We partition the data into distinct domains by academic program to emulate distributional differences. An Artificial Neural Network (ANN) model is first trained on a source domain and then fine-tuned on a target domain with a subset of layer weights frozen. We evaluate model performance using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and coefficient of determination (R^2), comparing the proposed transfer learning approach against a baseline without transfer. The results show that transfer learning significantly improves prediction accuracy: RMSE and MAE are reduced while R^2 increases on the target domain, indicating better generalization. The findings demonstrate that an ANN-based transfer learning method can effectively mitigate domain shift in student performance prediction. This study presents the benefits of transfer learning in an educational context by using attribute-based domain separation, offering a practical approach for academic institutions to predict student outcomes across different programs or semesters.