An Explainable Machine Learning–Based Educational Decision-Support Framework for the Early Identification of At-Risk Learners in E-Learning Environments: Cross-Dataset Validation Using OULAD and xAPI-Edu-Data

The early prediction of academic performance can support the timely identification of learners at risk of failure or withdrawal and enable targeted educational interventions. This study proposes an explainable machine learning–based educational decision-support framework for classifying learners’ academic outcomes in e-learning environments. The framework was evaluated using two independent educational datasets: the Open University Learning Analytics Dataset (OULAD) and xAPI-Edu-Data. For OULAD, seven source tables containing demographic, registration, assessment, and virtual learning environment interaction data were cleaned and integrated. Twenty-four predictive features were then derived from assessment performance, learning management system engagement, registration behavior, assessment type, and online activity type. Four classification algorithms—Random Forest, XGBoost, LightGBM, and CatBoost—were evaluated using an ۸۰:۲۰ stratified train–test split. RandomizedSearchCV was used to optimize the best-performing model, and SHAP was employed to explain model predictions. Feature engineering increased the Random Forest accuracy from ۶۳.۷۱% to ۶۷.۲۳%. The optimized XGBoost model achieved the best OULAD performance, with an accuracy of ۷۰.۶۹%, a weighted recall of ۰.۷۱, and a weighted F۱-score of ۰.۶۹. On xAPI-Edu-Data, XGBoost also achieved the highest accuracy at ۷۹.۱۷%. The findings indicate that XGBoost performed consistently across structurally different educational datasets, while the effectiveness of feature engineering depended on the richness and structure of the original attributes.

کلیدواژه ها:

academic performance prediction ، learning analytics ، explainable machine learning ، early-warning systems ، OULAD ، XGBoost