An Explainable Machine Learning Framework For Personalized Educational Intervention Using Learning Analytics And External Validation

The rapid expansion of Learning Management Systems (LMSs), e-learning platforms, and intelligent learning environments has generated massive volumes of educational data that can be leveraged to analyze learning behaviors, predict academic performance, and support data-driven decision-making in educational systems. Learning analytics enables the early identification of at-risk students, the design of targeted educational interventions, and the improvement of learning quality through the analysis of these data. Despite significant advances in machine learning algorithms, most previous studies have primarily focused on improving predictive accuracy while paying limited attention to model explainability, generalizability, and practical applicability in educational decision-making. Furthermore, the majority of existing studies evaluate predictive models using only a single dataset, with limited consideration given to external validation or the transformation of predictive outputs into actionable educational interventions. Learning analytics provides valuable opportunities to identify at-risk students, design targeted educational interventions, and improve learning outcomes through the systematic analysis of educational data. Nevertheless, although machine learning techniques have substantially enhanced predictive performance in this domain, previous studies have largely concentrated on prediction accuracy while overlooking model interpretability, educational applicability, and generalizability across different educational environments. In addition, external validation has rarely been incorporated, and predictive results have seldom been translated into practical educational decision-support strategies. This study proposes a comprehensive explainable machine learning framework for academic performance prediction and personalized educational intervention. The proposed framework utilizes demographic characteristics, academic background, assessment outcomes, and student interaction indicators extracted from the Open University Learning Analytics Dataset (OULAD). After feature engineering, these data are employed to train and evaluate four state-of-the-art machine learning algorithms, namely Random Forest, XGBoost, LightGBM, and CatBoost. To improve model transparency and identify the factors influencing predictions, SHAP (SHapley Additive explanations) is employed to interpret the selected model, after which students are categorized into different educational risk groups according to the most influential predictive features. Furthermore, the generalizability of the proposed framework is assessed through external validation using the xAPI-Edu-Data dataset. The experimental results demonstrate that CatBoost outperforms the other evaluated algorithms in predicting students' academic outcomes. SHAP-based analysis clearly identifies the most influential factors associated with academic success and failure, thereby facilitating the design of personalized educational interventions. The external validation results further confirm the satisfactory generalizability of the proposed framework across an independent educational dataset. Overall, by integrating predictive modeling, explainable artificial intelligence, educational risk profiling, and external validation, the proposed framework provides a reliable foundation for developing data-driven educational decision support systems and can assist instructors, academic advisors, and educational administrators in the early identification of at-risk students and the implementation of evidence-based personalized educational interventions.

کلیدواژه ها:

Learning Analytics ، Explainable Artificial Intelligence (XAI) ، SHAP ، Academic Performance Prediction ، CatBoost ، Personalized Educational Intervention