Despite the widespread use of deep learning models, particularly the Artificial Neural Network (ANN), for predicting student performance, their opaque nature often results in unreliable learned connections that deviate from educational expertise. This limitation hampers both reliability and future performance enhancements.
In this research project, an ANN was developed utilizing Shapley Additive Explanations (SHAP) and a public dataset of 395 Portuguese high school students’ math performance data. The investigation pinpointed critical factors affecting math performance and uncovered discrepancies between the ANN’s acquired relationships and established educational knowledge.
To tackle this challenge, we introduced the Student Performance Prediction Explanation (SPPE) algorithm to refine the ANN by reconsidering the importance of 30 features based on educational insights. Both global and local interpretability assessments were carried out to monitor changes in feature importance.
The study revealed a notable 26.9% enhancement in prediction accuracy for the refined ANN model aligned with educational expertise compared to its original version. Moreover, it surpassed various conventional machine learning algorithms.
Subsequent tests confirmed the adaptability of the SPPE strategy across different ANN structures, showcasing its resilience within this dataset and reinforcing its broad applicability and tangible benefits.
The study’s outcomes highlight how integrating educational expertise can enhance student performance forecasting, paving the way for transparent neural network frameworks and actionable guidance for diverse educational scenarios.