Abstract Predicting student dropout and enabling targeted interventions remain key challenges in modern education, specifically within online personalized learning environments. This study proposes an explainable machine learning framework for student dropout prediction and tailored interventions within online personalized education environments. Our emphasis is on three main areas: (i) evaluating the long-term stability of predictive accuracy and interpretability across several course settings, (ii) including student learning styles as contextual characteristics in the prediction model, and (iii) creating instance-level explanations to guide focused interventions. We trained ensemble models (Random Forest and XGBoost) on a publicly available dataset comprised of demographic characteristics, engagement patterns, and learning preferences which provided the basis for SHAP-based global and local interpretations. Our results show good and consistent performance throughout courses, most notably in Data Science (accuracy = 84%), with somewhat lower scores in Web Development. Although somewhat important in the decision process of the model, learning-style characteristics revealed little difference in dropout rates across categories, implying their predictive function might represent more general learner traits than causative influences. Instance-level SHAP explanations provide practical analysis of the particular elements causing each student's dropout risk, hence guiding the creation of tailored treatments. These results confirm the possibility of unified, explainable artificial intelligence models to assist dropout prediction in several online learning contexts. They also underline the need for adaptable, data-driven systems that strike a compromise between predicted accuracy and interpretability to guide fair and scalable educational decisions.
Nti et al. (Fri,) studied this question.