Predicting school performance using a combination of traditional and non-traditional education data from South Africa

The application of big data analytics in education is transforming learning, teaching and administration in schools. Current Education Data Mining (EDM) research focuses on teaching and personalized learning in higher institutions mostly in western countries with limited research conducted in African countries. Most research has been conducted using small datasets, simple learning analytics techniques and machine learning black box models to predict students’ performance. Black box modelling approaches use complex structures which are difficult to be easily interpreted by stakeholders. We synthesize EDM approaches and tree based machine learning techniques to identify important features that can predict school performance across African countries such as South Africa. We apply LightGBM a gradient boosting framework and interpretable tree based algorithms on combined data sources from community surveys, school master lists and examination results to perform feature importance. The challenge faced in EDM research is limited education data sources, we merged different existing datasets from government reports and archives. We used community survey data to determine the standards of living in secondary schools within those communities. Cell phone internet, toilets, security, usable water sources, number of teachers and students, school location, and family head were identified as control variables impacting the attainment of schools. LightGBM, underlies the developed prediction model. It empowered the model with high accuracy, stability and easy interpretation hence outperforming XGBoost, decision tree and random forest algorithms.