Predicting student academic outcomes is a critical task in Learning Analytics, yet the adoption of advanced predictive models is often hindered by their "black-box" nature. This experimental study evaluates the performance of three machine learning architectures—Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Long Short-Term Memory (LSTM) networks—on two benchmark educational datasets: xAPI-Edu-Data and the Open University Learning Analytics (OULA) dataset. Beyond mere predictive accuracy, we integrate post-hoc XAI techniques, specifically SHAP (SHapley Additive explanations) and LIME (Local Interpretable Model-agnostic Explanations), to quantify the contribution of specific behavioral and demographic features. Our results demonstrate that while XGBoost achieved the highest predictive accuracy (89.2%), the XAI layer revealed that "Resource Interaction Frequency" and "Parental Involvement" were the most significant predictors across all models, often overriding demographic factors such as nationality or gender. Furthermore, we observe that deep learning models (LSTM) provide superior temporal sensitivity but require XAI to decode the "moment of risk" during a semester. This paper provides empirical evidence that XAI can maintain high performance while offering the transparency required for effective pedagogical intervention, ultimately fostering a "Glass-Box" environment where AI serves as a collaborative consultant to the educator rather than an opaque judge. We conclude that XAI is a prerequisite for ethical AI deployment in sensitive educational environments, ensuring that algorithmic decisions are justifiable, transparent, and pedagogically sound.
G et al. (2026) studied this question.