Generating and Understanding Predictive Models for Student Attrition in Public Higher Education

Student attrition in higher education institutions is specially harmful in public funded institutions, and machine learning models have shown promising results predicting it and providing insightful information on its underlying causes. This work proposes using gradient boosting on decision trees to generate a high quality predictive model on data for almost 7,000 students of the University of Brasília, and applying SHAP to extract meaningful and comprehensive information on student attrition from it. Initial results indicate a strong predictive accuracy with over 90% recall and 85% f-score values and the SHAP analysis highlighted the importance of specific academic features, such as the student's performance on specific courses and all around performance during a period. More interestingly, the analysis provided insights on non academic features like age, commute time and whether the student enrolled through a social or racial quota or not. This may lead to better models and enhancing the discussions and changes on educational policies in the institution such as modifying a curricular structure or pursuing student housing closer to campuses.