Predicting National School Performance for Policy Making in South Africa

This paper presents an Education Data Mining (EDM) approach and machine learning techniques that were used to identify important features that can predict performance of high schools in South Africa. In order to be able to extract these factors we use interpretable machine learning algorithms to make it easier to translate the predictive power into actionable information that can be used by policymakers. Logistic regression and Light Gradient Boosting (LightGBM) and tree-based algorithms were applied on combined data sources from community surveys, school master lists and school government reports to perform feature importance and training of prediction models. Availability of clean water, toilets, hospitals, electricity, household goods, cellphone internet and safety in the communities were identified as important variables impacting the performance of schools. The two algorithms; LightGBM and Logistic regression, underlies the developed prediction models and empowered the models with high accuracy, stability, and easy interpretation as shown by the odd ratios and SHapley Additive exPlanations (SHAP) values.