Olameta/afrilearn-lens

XGBoost + SHAP model for student engagement detection in low-resource African learning environments using the OULAD dataset # AfriLearn Lens Behavioural engagement detection for students in low-resource African learning environments using the OULAD dataset. Link : AfriLearn Lens: Explainable… ## Overview AfriLearn Lens applies XGBoost and SHAP explainability to predict student engagement levels (Low, Medium, High) from passive Virtual Learning Environment (VLE) behavioural signals — without relying on demographic features. ## Key Finding SHAP analysis shows that behavioural features (active_days, unique_resources) are far stronger predictors of engagement than demographic features (gender, region, deprivation band) — challenging assumptions common in African education research. ## Model Performance - Accuracy: 80.78% - Best class: Low engagement (precision 0.92, recall 0.90) - Model: XGBoost classifier (200 estimators, max depth 6) ## Dataset OULAD — Open University Learning Analytics Dataset Source: Kaggle (anlgrbz/student-demographics-online-education-dataoulad) ## Features Used - active_days — number of unique days active on VLE - unique_resources — number of distinct resources accessed - resource_rate — exploration intensity (resources per active day) - consistency — regularity of access (active days per resource) - Demographics: gender, region, highest_education, imd_band, age_band, disability, num_of_prev_attempts, studied_credits ## Results ### Global Feature Importance (SHAP) ### Confusion Matrix ## Repository Structure - AfriLens_Code.ipynb — full pipeline notebook - shap_global_importance.png — SHAP bar chart - shap_class_high.png — SHAP beeswarm for High class - shap_class_low.png — SHAP beeswarm for Low class - shap_class_medium.png — SHAP beeswarm for Medium class - confusion_matrix.png — confusion matrix heatmap ## Author Abdussomad Olayiwola 400-level B.Tech Computer Science, LAUTECH AI and Data Science Instructor, Africa Research Center (ARC) GitHub: github.com LinkedIn: linkedin.com