
Predicting and understanding student engagement in online learning environments remains a critical challenge, particularly in African contexts where low-resource infrastructure limits the depth of available data. This paper introduces AfriLearn Lens, a machine learning pipeline that applies XGBoost classification and SHAP-based explainability to predict three-tier student engagement levels (Low, Medium, High) from Virtual Learning Environment (VLE) behavioural signals, using the Open University Learning Analytics Dataset (OULAD). The model achieves 80.78% accuracy on a held-out test set of 6,519 students. SHAP analysis reveals that temporal behavioural features — specifically active_days and unique_resources — are substantially stronger predictors of engagement than demographic variables. All code and results are available at github.com