Student Success Prediction in Digital Learning Environments

Abstract

Acknowledging the risk of perpetuating bias in AI-driven student success prediction, this study introduces a fairness-conscious machine learning framework that aims to balance predictive accuracy with ethical responsibility in digital learning settings. Using a dataset of 5,000 anonymized student records, three models, Random Forest (RF), Gradient Boosting (GBM), and Support Vector Machine (SVM), were developed to forecast academic outcomes. Model evaluation combined standard metrics (accuracy, precision, recall, and F1-score) with fairness measures such as demographic parity, equal opportunity, and disparate impact ratio to explore trade-offs between accuracy and fairness. Results indicated that while RF and GBM had slightly higher accuracy, SVM demonstrated more consistent fairness across demographic groups, emphasizing its stronger balance between predictive power and equity. A fairness-centered optimization method was applied to embed fairness constraints directly into model training, showing that both accuracy and fairness can be improved simultaneously rather than being in opposition. The framework integrates fairness throughout data preprocessing, model development, and post-prediction review, promoting transparent and responsible decision-making. By aligning with international ethical AI standards from UNESCO and the OECD, this research provides a practical pathway for creating educational prediction systems that enhance inclusion, minimize bias, and build trust in digital learning environments.