Student Academic Risk Prediction and Personalized Learning Recommendation Research Based on User Portraits of Learning Situation

Authors

  • Haochen Jiang Author

DOI:

https://doi.org/10.70693/rb8zpx45

Keywords:

academic risk early warning, educational data mining, XGBoost, SHAP, online learning

Abstract

Early identification of students with academic difficulties is challenging in traditional teaching scenarios. This study builds an academic-risk early-warning framework based on educational data mining on the Open University Learning Analytics Dataset (OULAD). After missing-value handling and min–max normalization of online learning records, key features such as learning duration, assignment completion rate and test scores are extracted to establish multi-dimensional student profiles, from which low-, medium- and high-risk labels are constructed. A classification model (Extreme Gradient Boosting (XGBoost) combined with the multi-dimensional profile features) is applied for academic-risk prediction, and the SHapley Additive exPlanations (SHAP) method is adopted to achieve model interpretability and mitigate the black-box defect of machine learning models. Comparative and ablation experiments show that the proposed model achieves an Accuracy of 0.831 and a macro-F1 of 0.808, exceeding the F1 scores of Logistic Regression and Random Forest by 0.095 and 0.051, respectively, and that each component (the profile module and the early-stopping mechanism) improves overall prediction performance. SHAP analysis quantifies how different learning features affect academic outcomes, helping teachers screen at-risk students and design personalized intervention strategies. Limited by the single-institution data source, the generalization ability of the model still needs to be verified on datasets from other universities. Future work will incorporate multi-source data, including classroom interaction, to refine the model and offer decision support for smart education.

Downloads

Published

2026-11-20

Issue

Section

Articles