Ensemble machine learning for predicting academic performance in STEM education

Science, Technology, Engineering, and Mathematics (STEM) education spans all levels, from preschool to postgraduate studies, fostering growth in social, cognitive, and psychomotor skills. In Ethiopia, the aim of STEM education is to cultivate engineers and scientists who can drive economic competitiveness on a global scale. Understanding and predicting student performance is vital for educators to identify areas for improvement and enhance academic outcomes. Unfortunately, Sub-Saharan Africa faces a significant challenge, with a low number of graduates in STEM fields. In countries like Ethiopia, students often struggle with science and mathematics. To tackle these issues, our research focused on developing a predictive model for STEM students using advanced ensemble machine learning algorithms. We gathered secondary data from the University of Gondar, Addis Ababa University, and Bahir Dar University, employing techniques such as Random Forest, CatBoost, Extreme Gradient Boosting, Gradient Boosting, Decision Trees, Logistic Regression, and Support Vector Machines. We utilized One-vs-Rest (OVR) and One-vs-One (OVO) class decomposition to transform multiclass labels into binary classifications. The model's effectiveness was assessed through metrics like accuracy, precision, recall, and F1 score. Key factors influencing the academic success of STEM students included their transcripts, departmental affiliation, socio-economic status, parental education and occupation, age, region, residency, internship experience, credit hours, entrance exam results, and admission types. Notably, the Gradient Boosting algorithm, using One-vs-Rest class decomposition, achieved impressive results: 93.35% accuracy, 92.69% precision, 93.14% recall, and an F1 score of 92.90%. These findings are crucial for policymakers aiming to enhance the STEM education landscape in Ethiopia. We also encourage future researchers to explore deep learning techniques alongside additional variables for even greater insights.